3D Spectroscopy in Astronomy PDF

This page intentionally left blank
3D SPECTROSCOPY IN ASTRONOMY
Simultaneously storing both spectral and spatial information, 3D spectroscopy offers a

new way to tackle astrophysical problems, and opens up new lines of research. Since
its inception in the eighties and early nineties, research in this field has grown enor-
mously. Large telescopes all around the world are now equipped with integral field units,
and two instruments of the James Webb Space Telescope will have integral field spectro-
scopic capabilities. Nowadays, more effort is dedicated to refining techniques for reducing,
analysing and interpreting the data obtained with 3D spectrographs.
Containing lectures from the seventeenth Winter School of the Canary Islands Astro-
physics Institute, this book explores new 3D spectroscopy techniques and data. A broad
and balanced presentation of research in this field, it introduces astronomers to a new
generation of instruments, widening the appeal of integral field spectroscopy and helping
it become a powerful tool in tackling astrophysical problems.
Canary Islands Winter School of Astrophysics
Volume XVII
Editor in Chief
F. Sánchez, Instituto de Astrofı́sica de Canarias
Volumes in this series

I. Solar Physics
II. Physical and Observational Cosmology
III. Star Formation in Stellar Systems
IV. Infrared Astronomy
V. The Formation of Galaxies
VI. The Structure of the Sun
VII. Instrumentation for Large Telescopes: a Course for Astronomers
VIII. Stellar Astrophysics for the Local Group: a First Step to the Universe
IX. Astrophysics with Large Databases in the Internet Age
X. Globular Clusters
XI. Galaxies at High Redshift
XII. Astrophysical Spectropolarimetry
XIII. Cosmochemistry: the Melting Pot of Elements
XIV. Dark Matter and Dark Energy in the Universe
XV. Payload and Mission Definition in Space Sciences
XVI. Extrasolar Planets
XVII. 3D Spectroscopy in Astronomy
XVIII. The Emission-Line Universe
Participants of the XVII Canary Islands Winter School of Astrophysics
3D SPECTROSCOPY IN
ASTRONOMY
XVII Canary Island Winter School of Astrophysics
Edited by
EVENCIO MEDIAVILLA, SANTIAGO ARRIBAS,
M A R T I N M . R O T H , J O R D I C E P A - N O G U É and
F R A N C I S C O S Á N C H E Z
Instituto de Astrofı́sica de Canarias, Tenerife
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,
São Paulo, Delhi, Dubai, Tokyo
Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521895415
© Cambridge University Press 2009
This publication is in copyright. Subject to statutory exception and to the

provision of relevant collective licensing agreements, no reproduction of any part
may take place without the written permission of Cambridge University Press.
First published in print format 2010
ISBN-13 978-0-511-76986-3 eBook (NetLibrary)

ISBN-13 978-0-521-89541-5 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy

of urls for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
Contents
List of contributors page viii
List of participants ix
Preface xi
Acknowledgements xii
List of abbreviations xiii
1 Introductory review and technical approaches 1

Martin M. Roth
2 Observational procedures and data reduction 40

James E.H. Turner
3 3D spectroscopic instrumentation 87
Matthew A. Bershady
4 Analysis of 3D data 126

Pierre Ferruit
5 Science motivation for integral field spectroscopy and Galactic studies 158
Frank Eisenhauer
6 Extragalactic studies and future integral field spectroscopy science 200

Luis Colina
7 Tutorials: How to handle 3D spectroscopy data 225

Sebastián F. Sánchez, Begoña Garcı́a-Lorenzo and Arlette Pécontal-Rousset
vii
Contributors
Matthew A. Bershady, University of Wisconsin, Madison, USA

Luis Colina, Departamento de Astrofı́sica, Molecular e Infrarroja, Instituto de Estruc-
tura de la Materia, CSIC, Spain
Frank Eisenhauer, Max-Planck-Institut für Extraterrestrische Physik, Germany
Pierre Ferruit, Observatoire de Lyon, CRAL, Université de Lyon, France
Begoña Garcı́a Lorenzo, Instituto de Astrofı́sica de Canarias, Spain
Arlette Pécontal-Rousset, Observatoire de Lyon, CRAL, Université de Lyon,
France
Martin M. Roth, Astrophysikalisches Institut, Potsdam, Germany
Sebastián F. Sánchez, Centro Astronómico Hispano-Alemán de Calar Alto, Spain
James E.H. Turner, Gemini Observatory, Chile
viii
Participants
Ahumada, Andrea Verónica Observatorio Astronómico de Córdoba

(Argentina)
Antichi, Jacopo Universita di Padova (Italy)
Brown, Warren Smithsonian Astrophysical Observatory
(USA)
Cardiel, Nicolas Centro Astronómico Hispano Alemán (Spain)
Cardwell, Andrew Open University/Isaac Newton Group
(UK/Spain)
Castillo-Morales, África Universidad Complutense de Madrid (Spain)
Cervantes Rodrı́guez, José L. Instituto de Astrofı́sica de Canarias (Spain)
Charcos Llorens, Miguel University of Florida (USA)
Christensen, Lise Astrophysikalisches Institut Postdam
(Germany)
de Hoyos Fernández de Cordova, Carlos Universidad Autónoma de Madrid (Spain)
de Lorenzo-Cáceres Rodrı́guez, Adriana Instituto de Astrofı́sica de Canarias (Spain)
Dı́az, Cristina Universidad Complutense de Madrid (Spain)
Dı́az, Rubén Joaquı́n Observatorio Astronómico de Córdoba
(Argentina)
Fernández Ontiveros, Juan A. Instituto de Astrofı́sica de Canarias (Spain)
Ferrand, Mickaël CNRS Lab. Sciences de l’Image, de
l’Informatique et de la Télédétection (France)
Ferrero, Patrizia Thueringer Landssternwarte Tautenburg
(Germany)
Flores-Estrella, Roberto Atahualpa Universidad Nacional Autónoma Mexico
(Mexico)
Floyd, David Space Telescope Science Institute (USA)
Fuentes Carrera, Isaura Inst. Astronomı́a, Geofı́sica y C.
Atmosféricas (Brazil)
Garcı́a-Benito, Rubén Universidad Autónoma de Madrid (Spain)
Garcı́a-Marı́n, Macarena Instituto de la Materia-CSIC (Spain)
Gavrilovic, Natasa Astronomical Observatory Belgrade, Serbia
and Montenegro
Germeroth, Andre University of Heidelberg (Germany)
Gieles, Mark Utrecht University (The Netherlands)
Gilli, Gabriela Instituto de Astrofı́sica de Canarias (Spain)
Gómez, Matı́as Universidad de Concepción (Chile)
Gómez-Cambronero Alvarez, Pedro GRANTECAN (Spain)
González Pérez, José M. Instituto de Astrofı́sica de Canarias (Spain)
Goodsall, Tim University of Oxford (UK)
Gössl, Claus A. Ludwig-Maximilians-Universitaet Muenchen
(Germany)
Gustafsson, Maiken University of Aarhus (Denmark)
Hernández Peralta, Humberto Instituto de Astrofı́sica de Canarias (Spain)
Herrmann, Kimberly Pennsylvania State University (USA)
Houghton, Ryan University of Oxford (UK)
ix
x Participants
Jahnke, Knud Max Planck Institute for Astronomy
(Germany)
Jalobeanu, André CNRS Lab. Sciences de l’Image, de
l’Informatique et de la Télédétection (France)
Kehrig, Carolina Instituto de Astrofı́sica de Andalucı́a (Spain)
Kelz, Andreas Astrophysikalisches Institut Postdam
(Germany)
Kjaer, Karina European Southern Observatory (Germany)
Koehler, Ralf Max-Planck-Institut für Extraterrestrische
Physik (Germany)
Kristensen, Egstrom Observatoire de Paris-Meudon (France)
Lagos Lizana, Patricio Andrés Observatorio Nacional ON/MCT (Brazil)
Lara López, Martiza Arlene Instituto de Astrofı́sica de Canarias (Spain)
López Martı́n, Dr. Luis Instituto de Astrofı́sica de Canarias (Spain)
Lorente Espin, Oscar Universidad Politécnica de Cataluña (Spain)
Lyubenova, Mariya European Southern Observatory (Germany)
Maier, Millicent University of Oxford (UK)
Martı́n Gordón, David Instituto de Astrofı́sica de Andalucı́a (Spain)
Moiseev, Alexei Special Astrophysical Observatory (Russia)
Monreal Ibero, Ana Astrophysikalisches Institut Postdam
(Germany)
Mora, Marcelo European Southern Observatory (Germany)
Nowak, Nina Max-Planck-Ins Extraterrestrische Physik
(Germany)
Ocvirk, Pierre University of Central Lancashire (UK)
Parker, Richard John University of Sheffield/ING (UK/Spain)
Pastorini, Guia Universita di Florence (Italy)
Prez Gallego, Jorge University of Florida (USA)
Prez-Montero, Enrique Universidad Autónoma de Madrid (Spain)
Povic, Mirjana Instituto de Astrofı́sica de Canarias (Spain)
Ragaini, Silvia Universita di Padova (Italy)
Ramirez Ballinas, Isidro Universidad Nacional Autónoma Mexico
(Mexico)
Rebaza Castillo, Ovidio Instituto de Astrofı́sica de Andalucı́a (Spain)
Rix, Samantha Anne Isaac Newton Group Telescopes (Spain)
Scheepmaker, Remco Astronomical Institute Utrecht (The
Netherlands)
Smirnova, Aleksandrina Special Astrophysical Observatory (Russia)
Stoklasova, Ivana Observatoire de Lyon (France)
Tamburro, Domenico Max-Planck-Institute for Astronomy
(Germany)
Trachternach, Clemens Ruhr-Universität Bochum (Germany)
Valdivielso Casas, Luisa Instituto de Astrofı́sica de Canarias (Spain)
van Eymeren, Janine Ruhr-Universität Bochum (Germany)
Vidrih, Simon University of Cambridge (UK)
Wang, Wei Peking University (China)
Preface
3D spectroscopy has a relatively short history. Most of the present instrument concepts
were developed in the 1980s and early 1990s. During those pioneering years a great deal of
work was done in optical labs in an attempt to understand how the optical fibres, micro-
lenses and image slicers behave. Only a few groups (often formed by one or two people)
worked on this topic. Communications were not very good, which explains why virtually
all the groups decided to refer to this technique by a different name. So we ended up
with ‘spectral imaging’, ‘bidimensional spectroscopy’, ‘integral field spectroscopy’, ‘two-
dimensional spectroscopy’, ‘3D spectroscopy’, etc.
During those years it was more than doubtful whether this technique was going to
be useful at all. In fact, it looked like a kind of curiosity of limited practical interest
to astronomy. However, in the 1990s the first scientific results were obtained and they
inmediately produced a change of perception.
In the last few years investment in this type of instrumentation has been enormous.
Large telescopes all around the world are now equipped with integral field units. Two
instruments of the future James Webb Space Telescope will also have integral field spec-
troscopic capabilities, etc. Instead of being based in the optical lab trying to characterize
optical fibres or micro-lenses, more effort is dedicated nowadays to refining techniques
for reducing, analysing and interpreting the data obtained with a new generation of 3D
spectrographs. Clearly, we are in a wholly different phase; by attending the lectures and
viewing the posters of our Winter School, it is clear that 3D spectroscopy has truly
arrived on the scene.
This Winter School is particularly timely. It is aimed at instructing a new generation
of astronomers in a new generation of instruments. This will probably be the first and
last school on this topic. In a few years’ time, organizing a similar school would seem like
organizing a school today on ‘imaging’. However, the string of recent scientific conferences
on 3D spectroscopy and the over-subscription to this school would seem to indicate that
this is the right topic at this time.
The Instituto de Astrofı́sica de Canarias (IAC) is specially pleased to be organizing a
Winter School on the topic of 3D spectroscopy because it has contributed to this field
since its inception. More than 15 years ago a small group of researchers at the IAC started
working in this field. They set up from scratch an optical fibre lab at the institute and
developed the first experimental integral field systems for the telescopes on La Palma.
That was just the beginning of the IAC’s contribution to the development and diffusion of
this technique, a contribution that continues today to extend on several fronts, including
the organization of this workshop in collaboration with the European Euro3D network.
The aim of this Winter School is to widen the appeal of integral field spectroscopy
beyond the limited community of experts and to help it become a powerful tool in the
hands of a new generation of astronomers for tackling new (and old) astrophysical prob-
lems. We thank all the participants, lecturers and students for their valuable contributions
to this objective.
The Editors
xi
Acknowledgements
The editors want to express their warmest gratitude to the lecturers for their efforts
in preparing their classes and the chapters of this book. We also wish to thank our
efficient secretaries: Nieves Villoslada and Lourdes Gonzalez, and many staff members
of the IAC: Jess Burgos, Begoña López Betancor, Carmen del Puerto, Ramón Castro,
Terry Mahoney, Miguel Briganti and the technicians of Servicios Informaticos Comunes
del IAC. We acknowledge Ismael M. Delgado for the technical edition of this book for
Cambridge University Press.
Finally, we also wish to thank the Astrophysical Institute Potsdam and the European
Research Training Network Euro3D for their support to the finances, and the organization
and the following institutions for their economical support: The European Commission,
the Spanish Ministry of Education and Science, Iberia, the local governments (cabildos)
of the islands of Tenerife and La Palma, and the Puerto de La Cruz Town Council.
xii
Abbreviations
2dF The Two Degree Field system

2D-FIS 2-Dimensional Fiber ISIS System
AAT Anglo-Australian Telescope
ACS The Advance Camera for Surveys
AOB Adaptive Optics Bonnette
APO Apache Point Observatory
APOGEE Apache Point Observatory Galactic Evolution Experiment
ARGUS
ARIES Arizona infrared imager and echelle spectrograph
CAHA Centro Astronómico Hispano Alemán
CCD charge coupled device
CFHT Canada–France–Hawaii Telescope
CIGALE Scanning Perot–Fabry Interferometer, comparable with the UK
TAURUS system, developed by Marseille Observatory to work
at the Cassegrain focus of the 3.6 m CFH Telescope
CIRPASS Cambridge Infrared Panoramic Survey Spectrograph
COHSI Cambridge OH Suppression Instrument
EFOSC ESO Faint Object Spectrograph and Camera
ESI Echellette Spectrograph and Imager
ESO European Southern Observatory
F2T2 Flamingos-2 Tandem Tunable Filter
FaNTOmM Fabry–Perot of New Technology for the Observatoire du mont
Megantic
FGS-TF Tunable Filter
FISICA The Florida Image Slicer for Infrared Cosmology and
Astrophysics
FITS Flexible Image Transport System
FLAMES Fibre Large Array Multi Element Spectrograph
FLAMES/GIRAFFE FLAMES Intermediate Resolution Spectrometer
FLAMINGOS The Facility Near-Infrared Wide-field Imager and Multi-Object
Spectrograph for Gemini
FORS Focal Reducer and low dispersion Spectrograph
FWHM full-width-at-half-maximum
GMOS Gemini Multi-Object Spectrograph
GNIRS Gemini Near Infrared Spectrograph
GriF Grating Infrared Fabry–Perot
GTC Gran Telescopio CANARIAS
HARPS High Accuracy Radial Velocity Planet Searcher
HET Hobby–Eberly Telescope
HEXAFLEX Hexagonal Lattice Fiber Linked Experimental
HEXAFLEX-II Hexagonal Lattice Fiber Linked Experimental II
HIFI Hawaii Imaging Fabry–Perot Interferometer
HIRES High Resolution Echelle Spectrometer
HST Hubble Space Telescope
xiii
xiv Abbreviations
IAU International Astronomical Union
ICA Instituto de Astrofı́sica de Canarias
IDL Interactive Data Language
IFS Integral Field Spectroscopy
IFU Integral Field Unit
IMACS-IFU Inamori Magellan Cassegrain Spectrograph–Integral Field Unit
IRAF Image Reduction and Analysis Facility
ISIS Intermediate-Dispersion Spectroscopic and Imaging System
JWST James Webb Space Telescope
KMOS K-Band Multi-Object Spectrometer
KPNO Kitt Peak National Observatory
MACAO Multi-Application Curvature Adaptive Optics
MCAO Multi-Conjugate Adaptive Optics
MIDAS Mid-Infrared Asteroid Spectroscopy
MIRI Mid-InfraRed Instrument
MOS Multi-Object Imaging Spectrograph
MPE 3D Max Planck Institut für Extraterrestrische Physik 3D
MPFS Multi-Pupil Field Spectrograph
MUSE Multi-Unit Spectroscopic Explorer
NaCo short for NAOS-CONICA (Nasmyth Adaptive Optics System –
Near-Infrared Imager and Spectrograph)
NIC-FPS Near-Infrared Camera and Fabry–Perot Spectrometer
NICMOS Near Infrared Camera and Multi-Object Spectrometer
NIFS Near-Infrared Integral Field Spectrograph
NIRSpec Near-Infrared Spectrograph
NOAO National Optical Astronomy Observatory
NOT Nordic Optical Telescope
OAN-SPM National Astronomical Observatory at San Pedro Martir
OASIS Optically Adaptive System for Imaging Spectroscopy
OHP Observatoire de Haute-Provence
OMM Observatoire du mont Megantic
ORM Observatorio del Roque de los Muchachos
OSIRIS Optical System for Imaging and low/intermediate-Resolution
Integrated Spectroscopy
PACS Photoconductor Array Camera and Spectrograph
PMAS Potsdam Multiaperture Spectrophotometer
PPAK PMAS fiber pack
PSF Point Spread Function
PUMA UNAM (Universidad Nacional Autónoma de Mexico) Scanning
Fabry–Perot Interferometer
PUMILA Near-infrared Spectrograph for the Kinematic Study of the
Interstellar Medium
PYTHEAS Prisme Interferomètre Trames de lentilles pour l’Holometrie, et
l’Endoscopie des Astres et des Sources
QSO quasi-stellar object
Abbreviations xv
RFP Rutgers Fabry–Perot
RSS Robert Stobie Spectograph
RSS Row Stacked Spectra
RTN Research Training Network
SALT Southern African Large Telescope
SAO Smithsonian Astronomical Observatory
SAURON Spectrographic Areal Unit for Research on Optical Nebulae
SHM shared memory server
SILFID Spectrographe Integral Linarisation par Fibres de l’Image
Directe
SINFONI Spectrograph for Integral Field Observations in the Near
Infrared
SIS Stabilized Imager and Spectrometer
SMM Submillimetre galaxy
SNIFS SuperNovae Integral Field Spectrograph
SPIFFI Spectrometer for Infrared Faint Field Imaging
SWIFT Short Wavelength Integral Field Spectrograph
TF Tully–Fisher
TIGER Traitement Intégral des Galaxies par l’Etude de leurs Raies
TTF TAURUS Tunable Filter
UFTI UKIRT Fast-Track Imager
UIST UKIRT Imager-Spectrometer
UKIRT United Kingdom Infrared Telescope
VIMOS Visible Imaging Multi-Object Spectrograph
VIRUS Visible Integral-Field Replicable Unit Spectrograph
VLT Very Large Telescope
VPH Volume Phase Holographic
WFPC Wide-Field Planetary Camera
WHAM Wisconsin H-alpha Mapper
WHT William Herschel Telescope
WiFeS Wide-Field Spectrograph
WIYN Wisconsin–Indiana–Yale
WYFOS Wide-Field Optic Spectrograph
1. Introductory review and
technical approaches
MARTIN M. ROTH
1.1 Preface
The topic of the XVII IAC Winter School is ‘3D Spectroscopy’: a powerful astronomical
observing technique, which has been in use since the early stages of the first prototype
instruments about a quarter of a century ago. However, this technique is still not con-
sidered a standard common user tool among most present-day astronomers.
3D Spectroscopy (hereafter ‘3D’) is also called ‘integral field spectroscopy’ (IFS),
sometimes ‘two-dimensional’ or even ‘area’ spectroscopy, and commonly also ‘three-
dimensional’ spectroscopy; in other areas outside astronomy it is called ‘hyperspectral
imaging’, and so forth. It is already this diversity in the nomenclature that perhaps
reflects the level of confusion. For practical reasons, the organizers of this Winter School
and the Euro3D network (which will be introduced below) have adopted the terminology
‘3D’, which is intuitively descriptive, but, as a caveat early on, is conceptually mislead-
ing if we restrict our imagination to the popular picture of the ‘datacube’ (Figure 1.1).
Although this term will commonly be used throughout this book, we need to point out
for the reasons given later in the first chapter that the idealized picture of an orthogonal
cube with two spatial, and one wavelength, coordinate(s) is inappropriate in the most
general case.
Whatever the terminology, it is the aim of this Winter School to help alleviate the
apparent lack of insight into 3D instrumentation, its use for astronomical observations,
the complex problems of data reduction and analysis, and to spread knowledge among
a significant number of international young researchers at the beginning of their career.
The training of early stage researchers has always been the rationale in the tradition of
the well-established series of Instituto de Astrofı́sica de Canarias (IAC) Winter Schools,
but this is the first time in which a consortium of 11 leading European research institutes
appears as co-organizer, represented by the ‘Euro3D’ Research Training Network (RTN) –
which may well be recognized as the significant impact expected from this school. Euro3D
was funded by the European Commission (EC) from July 2002 to December 2005 under
! Astrophysikalisches Institut Potsdam (Germany, coordinator)

Framework Programme 5 (Walsh and Roth, 2002). The RTN participants are as follows:
! IoA University of Cambridge (UK)

! Durham University (UK)
! MPE Garching (Germany)
! European Southern Observatory (International Organization)
! Sterrewacht Leiden (The Netherlands)
! Observatoire de Lyon (France)
! CNRS Marseille (France)
! IFCTR Milano (Italy)
! Observatoire de Paris, Meudon (France)
! IAC Tenerife (Spain).
At the time of planning the RTN in 2001, these groups represented practically the entire
expertise in 3D instrumentation in Europe. As can be seen from a compilation of exist-
ing instruments and those under development at the time (Table 1.1), significant capital
1
2 Martin M. Roth
λ
Monochromatic maps
x
y
Each data cube slice represents

a monochromatic image
Visualization
... final data cube
Single or co-added spectra
Figure 1.1. Schematic representation of a datacube, produced as the result of 3D observations.

It can be interpreted as a stack of quasi-monochromatic images or, alternatively, as an assembly
of n × m spectra.
Figure 1.2. Participants of the Euro3D Kickoff Meeting, held at IAC, Tenerife, 2–4 July 2002.
investments have been realized at almost all major observatories, with a clear technolog-
ical lead in Europe. Disappointingly, however, the expertise within the user community
for the reception of the data from such instruments was found to be very limited. It was
confined almost entirely to members of the groups that have built the 3D instruments.
Therefore, and in response to the EC requirement to demonstrate ‘the scientific, tech-
nological, or socio-economic reasons for carrying out research in the field covered by the
! to promote 3D technique with selected science projects

network’, the project objectives of the Euro3D consortium were identified as follows:
! to develop new observing and data reduction techniques

! to provide training and education of students and young researchers
! to popularize 3D technique in various research fields
! to provide user support and general information
Table 1.1. Compilation of 3D instrumentation on some major telescopes (as of 2001). Notes: Bold face: 3D instruments built by
RTN members. (•): facilities accessible through Euro3D. 1 ) : mobile instrument.
Diameter Access 3D ∆λ Maximum FoV Maximum resolution

Telescope (m) Euro3D Instrument (µm) arcsec2 λ/∆λ Status
NGST 8.0 • IFMOS proposed

HST 2.4 • –
Keck 2 × 10 OSIRIS 0.9–2.5 3×6 4000 planned
GTC 10 • ATLANTIS 1–2.4 8×8 5000 planned
HET 10 –
LBT 2 × 8.4 • PMAS 1 ) 0.35–1.0 16 × 16 2200 private
SUBARU 8.3 Kyoto-3D
GEMINI 2 × 8.1 • GMOS 0.4–1.0 50 × 450 10000 common user
• GNIRS 1–5 20 × 20 5000 common user
• CIRPASS 0.9–1.8 35 arcsec2 3000 common user
• NIFS 1–2.5 3×3 5300 planned
VLT 4 × 8.1 • VIMOS 0.37–1 54 × 54 2000 common user
• NIRMOS 1–1.8 28 × 28 2500 common user
• SINFONI 1–2.5 8×8 4500 common user
• FLAMES 0.37–0.9 11.5 × 7.3 25000 common user
Magellan 2 × 6.5 • IMACS 0.4–0.9 50 × 50 5000 common user
Selentchuk 6.0 • MPFS 0.39–0.9 16 × 16 3000 common user
Hale 5.0 PIFS 1–5 5.4 × 9.6 3000 experimental
WHT 4.2 • INTEGRAL 0.39–1.0 1800 common user
• SAURON 0.47–0.54 39.5 × 32.9 2300 private
• TEIFU 0.47–0.54 4.1 × 5.9 private
CTIO 4.0 –
KPNO 4.0 –
AAT 3.9 • SPIRAL 0.48–1.0 22 × 11 11000 ‘expert’ user
CFHT 3.6 • TIGER 0.4–1.0 12 × 9 2000 decommissioned
• OASIS 0.4–1.0 15 × 12 3000 common user
ESO 3.6 m 3.6 • –
WYIN 3.5 DENSEPAK 0.38–1.0 90 fibers 22000 common user
Calar Alto 3.5 • PMAS 1 ) 0.35–1.0 16 × 16 2200 private
NTT 3.5 • –
TNG 3.5 • GOHSS 0.8–1.8 3000 planned
• MPE-3D 1 ) 1.5–2.2 6.4 × 6.4 2100 private
4 Martin M. Roth
Figure 1.3. The conceptually problematic picture of a ‘datacube’
! to provide feedback from observers to technology

! to identify technology transfer opportunities.
Organizing the XVII IAC Winter School is an important contribution to its major goals
and appropriately concludes the activities at the end of the lifetime of the RTN. It is
hoped that the participants of the Winter School will benefit from the collective expertise
presented in the lectures for the rest of their future scientific careers; furthermore that 3D
Spectroscopy shall experience more widespread use in the community to further exploit
and advance its unique potential for astronomy.
1.2 Introductory review

1.2.1 Conceptual outline
Principle of operation and terminology
We define integral field (3D) spectroscopy (IFS) as an astronomical observing method
that creates in a single exposure spectra of (typically many) spatial elements (‘spaxels’)
simultaneously over a two-dimensional field-of-view (FoV) on the sky. Owing to this
sampling method, each spaxel can be associated with its individual spectrum. Once
all of the spectra have been extracted from the detector frame in the process of data
reduction, it is possible to reconstruct maps at arbitrary wavelengths. For instruments
with an orthonormal spatial sampling geometry, the spectra can be arranged on the
computer to form a three-dimensional array, which is most commonly called a ‘datacube’.
Datacubes are also well known as the natural data product in radio astronomy. Note,
however, that there are many integral field spectrographs which do not sample the sky
in an orthonormal system. In this case the term datacube is misleading. We shall see
further below that atmospheric effects in the optical also make the term inappropriate
in the most general case.
Historically, various designations have been invoked to describe integral field spec-
troscopy (area/imaging/2D/3D/three-dimensional spectroscopy, etc.). As spectroscopy
involves one dimension per se (the wavelength scale), two-dimensional spectroscopy would
arguably also describe the situation correctly. However, the datacube picture and the per-
haps more explicit label ‘3D’ have conventionally led to the now accepted terminology
of three-dimensional spectroscopy, which is adopted throughout this book.
Instruments that create three-dimensional datasets in the sense described above, how-
ever not simultaneously but rather in some process of sequential data acquisition (scan-
ning), e.g. tuneable filter (Fabry–Perot) instruments, scanning long slits, etc., are not
strictly 3D spectrographs according to our definition. Because of similarities in the
Introductory review and technical approaches 5
Figure 1.4. The three major principles of operation of present-day IFUs (Source: J. Allington-
Smith).
RECTANGLE
SQUARE
PIXSIZE2
PIXSIZE1
PIXSIZE1
HEXAGON CIRCLE
PIXSIZE1 PIXSIZE1
Figure 1.5. Different types of spaxel geometries: square, rectangular, hexagonal, bare fiber
(circular).
analysis and interpretation of their data, however, such instruments will also be men-
tioned briefly below.
Methods of image dissection: spatial sampling

Integral field spectrographs have been built based on different methods of dissecting the
FoV into spaxels, e.g. optical fiber bundles, lens arrays, optical fibers coupled to lens
arrays, or slicers (Figure 1.4). We will discuss these methods further below.
The term ‘spaxel’ was introduced by the Euro3D consortium in order to distinguish
spatial elements in the image plane of the telescope from pixels, which are the spatial ele-
ments in the image plane of the detector (Kissler-Patig et al., 2004). The optical elements
that accomplish the sampling of the sky are often called ‘integral field units’ (IFUs), and
thus IFS is also sometimes called ‘IFU spectroscopy’. Spaxels can have different shapes
and sizes, depending on instrumental details and the type of IFU (Figure 1.5).
It is necessary to point out that, contrary to the persuasive implication of the datacube
picture, IFUs do not necessarily sample the sky on a regular grid, e.g. fiber bundles, where
6 Martin M. Roth
Figure 1.6. Field distortion in the fore-optics of an IFU affects the spatial sampling.
100
-enhanced AR coating
90 Broadband AR coating
Front-illuminated
80
70
60
E( )
50
Measured
40
30
20
10
0
300 400 500 600 700 800 900 1000 1100
Wavelength (nm)
Figure 1.7. Fairchild CCD 4K × 4K, 15 µm pixels (left). The QE of modern CCDs can be made
quite high, very nearly approaching 100% (right).
due to the manufacturing process individual fibers cannot be arranged to arbitrary pre-
cision (for a real fiber bundle IFU, see Figure 1.23). This is another reason why the
popular description of IFS as ‘like a CCD [charge coupled device] with a spectrum in
each pixel’ is misleading. Even if the manufacture of an IFU allows a perfectly regular
sampling pattern to be created, e.g. in the case of a hexagonal lens array, the sam-
pling is not necessarily orthonormal. Moreover, real optical systems create aberrations
and, sometimes, non-negligible field distortions (Figure 1.6), which also means that the
set of spectra extracted from the detector does not sample an orthogonal FoV on the
sky. Furthermore, the sampling method may be contiguous (e.g. lens array) with a fill
factor very close to unity, or non-contiguous (e.g. fiber bundle) with a fill factor of sig-
nificantly less than 100%. In all of these cases, it is possible to reconstruct maps at a
given wavelength through a process of interpolation and, repeating this procedure over
all wavelengths, to convert the result into a datacube. Note, however, that interpolation
often produces artifacts and generally involves loss of information.
Detectors and spectrographs

In order to better understand the potential and natural constraints of 3D instrumenta-
tion, it is useful to briefly review detectors, since they ultimately limit the total number of
spatial and spectral resolution elements, hence FoV and wavelength coverage. Depending
on the type of application, one would normally try to maximize the number of detector
pixels in order to maximize one or both of these parameters. A selection of modern large
area detectors is illustrated below. For visual wavelengths in the interval 0.35−1.0 µm,
CCDs are nowadays the most common detectors, featuring high quantum efficiency (QE)
and low read-out noise. Single chips with pixel resolutions of 4096 × 4096 (Figure 1.7)
presently represent the state-of-the-art, although much larger CCDs are under develop-
ment. In order to obtain even larger pixel numbers, buttable CCDs can be arranged to
form mosaic configurations similar to modern wide-field imaging cameras (MegaCAM,
OmegaCam, etc.), e.g. the detector for the IMACS IFU (Inamori Magellan Cassegrain
Spectrograph – Integral Field Unit), which is built as a mosaic of 2 × 3 2K × 4K CCD
chips. Image-intensified photon counting detectors that were still common in the 1980s
Figure 1.8. Rockwell 2K × 2K Infrared Array (left), 4K × 4K mosaic (right).
are no longer popular, except for read-noise limited applications with very low photon
count rates. An interesting new development is the photon-counting LLCCD, which uses
an on-chip avalanche gain technology for low noise at high frame rates (Jerram et al.,
2001).
In the near infrared (NIR) wavelength regime there has been enormous progress in
detector technologies, providing large area infrared (IR) arrays with pixel resolutions of
up to 2K × 2K and excellent electro-optical performance, and even larger arrays being
under development. A mosaic of 2 × 2 such buttable arrays is shown in Figure 1.8. While
a strong driver for this evolution has come from the development of the next generation
space telescope – the James Web Space Telescope (JWST) – it is also the demand for
detectors from NIR instrumentation for ground-based astronomy that has made NIR
astronomy a rapidly growing field.
The number of spectra delivered by present-day 3D spectrographs is typically in the
range of several hundreds up to a few thousand: Visible Imaging Multi-Object Spectro-
graph (VIMOS) IFU: 6400 (Le Fevre et al., 2003). The corresponding number of spaxels is
disappointingly small if one compares with direct imaging detectors: even the simple first
generation commercial CCD imagers were offering, for example, 384 × 576 ≈ 221 Kpixels:
a factor of 34 larger than the number of spaxels of the largest IFU to date. So what is
it that limits the size of an integral field spectrograph in terms of number of spatial and
spectral elements?
First of all, IFS is obviously more complex than direct imaging: for any spatial element,
a full spectrum must be accommodated. If n, m, k are the number of spatial resolution
elements in x and y, and of the spectral resolution elements, respectively, then the total
number N of required detector pixels is N = n × m × k. In comparison with direct imaging
where k = 1, IFS requires at least a factor of k more detector pixels. Clearly, instrumen-
tation for IFS is expensive in terms of detector space.
Another important factor is the spectrograph optical system. The input focal ratio as
determined by the telescope and any fore-optics, the magnification from the spectrograph
slit to the detector pixel size, as well as the total length of slit determine the spectrograph
parallel beam size and thus the overall size of the optical system.
For a large number of spectra, a wide-field optical system is typically required. Such
optical systems are challenging in terms of technical feasibility (controlling optical aber-
rations), and cost. As an example, Figure 1.9 shows the layout of the PMAS (Potsdam
Multiaperture Spectrophotometer) spectrograph optics, which is optimized to cover the
wavelength range 350−900 nm (Roth et al., 2005). It is worthwhile noting that the lens
sizes are 200 mm in diameter. While this size represents the state-of-the-art for modern
8 Martin M. Roth
Figure 1.9. PMAS fiber spectrograph as an example for an IFS optical system. Left: a cross
section, showing the layout of collimator, grating and camera. The parallel beam measures
150 mm in diameter. The magnified detail to the lower left illustrates the fiber interface. Right:
photograph of the camera, which has a total mass of 60 kg.
Figure 1.10. Different ways of allocating spectra on a detector: a) maximizing number of

spaxels with full wavelength coverage, b) maximizing number of spectral pixels, c) maximizing
number of spaxels with several banks of spectra, d) diagonal, maximizing number of spaxels at
expense of wavelength coverage.
refractive optical systems for spectroscopy, the apochromatic correction from the ultra-
violet (UV) to the NIR is unusual and quite demanding. The cost of such a system is
high, mainly because of the use of expensive materials, e.g. CaF2 , which are required for
high throughput and the large wavelength coverage. Instruments that are not designed
for general-purpose operation but rather for a specific science case can often be built
simpler and with less effort.
Given the total number of spaxels of the IFU, the critical question of how to allocate
the corresponding spectra on the detector arises. Obviously, the total number of detector
pixels is related to the number of spatial and spectral resolution elements like:
ΣPixels ≥ NSpaxels × NSpectra (1.1)
In most conventional 3D spectrograph designs the detector space is not contiguously
filled with spectra, basically in order to avoid cross-talk between adjacent spectra and
to facilitate data reduction. Novel slicer designs like, for example, SINFONI (Eisenhauer
et al., 2003) or the OSIRIS lens array IFU (Larkin et al., 2006) have, however, demon-
strated that this is not a necessary condition. Figure 1.10 shows several basic ways of
arranging the spectra. Generally, a trade-off needs to be made between maximizing the
number of spaxels, or the number of spectral pixels.
Figure 1.11. Typical CCD frame of an arc lamp exposure from a fiber bundle IFU. The direc-
tion of dispersion is horizontal. The magnified insert demonstrates how the instrumental profile
is given by the FWHM (full-width-at-half-maximum) of the projected fiber size on the detector
( ≈ 4 pixels).
a) b)
Figure 1.12. Examples for instrumental efficiency as a function of wavelength for different
gratings. a) PMAS with lens array IFU, b) PMAS with PPak IFU.
Spectral resolution and wavelength coverage

The spectral resolution in most 3D instruments does not reach the diffraction limit. It
is rather determined by the projected (pseudo-)slit width on the detector. Figure 1.11
illustrates this situation with an example dataset, which was obtained with a fiber-based
3D spectrograph.
As mentioned before, the wavelength coverage is normally tuned to the application and
scientific scope of an instrument. Two extreme examples on 4 m class telescopes would
! SAURON (WHT) [Spectrographic Areal Unit for Research on Optical Nebulae

be:
(William Herschel Telescope)], which has a wavelength coverage of 54 nm in the
! PMAS (CAHA 3.5 m) [Potsdam Multiaperture Spectrophotometer], with a wave-

operational wavelength interval of 450−1100 nm; and
length coverage of up to 680 nm in the interval of 350−900 nm. Instrumental efficiency

curves for the latter instrument are given in Figure 1.12.
Coupling the IFU to the spectrograph

Although this is a somewhat technical detail, it is worth pointing out the fundamental
difference between two different methods of coupling the IFU to the spectrograph optical
system. One method is the one that has conventionally been used for normal slit-based
spectrographs, namely to image the (dispersed) telescope focal plane (= slit) onto the
detector. This technique is normally used for the ‘image slicer’ type of 3D spectrograph.
10 Martin M. Roth
Table 1.2. Examples for 3 types of 3D spectrographs
Nspectra FoV [′′ ] scale [′′ ] R λ range/∆λ [nm]
SAURON LR 1577 41 × 33 0.94 1200 450–700/54

SAURON HR 1577 11 × 9 0.27 1400 450–700/54
VIMOS LR 6400 54 × 54 0.67 220 370–1000/400
VIMOS HR 1600 13 × 13 0.33 2700 420–870/240
8×8 0.25
SINFONI 2048 3×3 0.10 1500–4000 1050–2450
0.8 × 0.8 0.025
Another method has emerged with the introduction of optical fibers in that the image
information at the telescope focal plane is scrambled, and thus the one-to-one corre-
spondence between points on the image plane and on the detector plane is lost. This
principle is employed in an even more rigorous way in 3D spectrographs of the TIGER-
type (TIGER = Traitement Intégral des Galaxies par l’Etude de leurs Raies; named
after the prototype; Bacon et al., 1995), where a simple fore-optics system followed by a
lens array creates a tiny image of the entrance pupil (telescope aperture) for each lens.
These so-called micropupils are designed to typically measure a few tens of micrometers
in diameter. When fed to the collimator input plane of an imaging spectrograph, the fam-
ily of micropupils practically operates like the apertures of a conventional multi-object
spectrograph (e.g. the EFOSC-type), with the one significant distinction, namely that it
is the telescope pupil rather than the telescope image that forms the multi-slit configu-
ration. The important consequence is that the light distribution of the slit illumination
no longer matters in terms of wavelength definition, which is an important argument for
precision spectroscopy. The principle is not entirely new, but was first employed using
a ‘Fabry lens’ for photoelectric photometers in order to eliminate the photometric noise
introduced by seeing (Michlovic, 1972). Another variant of the TIGER micropupil tech-
nique is employed in fiber-optical 3D spectrographs, where the fiber bundle is connected
to a lens array.
Examples of real IFUs

For the purpose of making the techniques introduced in the preceding sections more
imaginable, three instruments are presented, which are meant to be representative for
the (1) fiber-optical, (2) micropupil lens array (TIGER), and (3) slicer types of 3D
spectrographs, namely the VIMOS (Visible Imaging Multi-Object Spectrograph; Le Fevre
et al., 2003), SAURON (Bacon et al., 2001), and SINFONI instruments (Spectrograph for
Integral Field Observations in the Near Infrared; Eisenhauer et al., 2005). Some important
parameters (number of spectra, FoV, spaxel size, spectral resolving power, wavelength
coverage) are listed in Table 1.2 (LR/HR stand for low resolution and high resolution
modes, respectively). Figure 1.13 shows examples of what the spectra look like on raw
CCD frames for these three instruments. The VIMOS frame shows a science exposure
from a stellar field, so some spectra are illuminated well, while most others contain only
sky. Note that the direction of dispersion is oriented horizontally, and that the frame
contains 4 banks of spectra that are placed as columns next to each other, with a gap
between bank numbers 1 and 2. Note also that there are several artifacts as a result
from zero and higher order contamination. The SAURON example shows a continuum
flat-field exposure, which exhibits the characteristic pattern of the diagonal arrangement
of the total 1577 spectra, whose density is so high that individual spectra are difficult to
a)
b) c)
Figure 1.13. a) Raw detector frames from VIMOS IFU; b) SAURON; c) SPIFFI. See text for
explanation.
discern. Finally, the SINFONI example illustrates the appearance of slicer data, which
at first glance looks like noise. Closer inspection reveals the staggered slit arrangement
as a vertical pattern right from the center. It is obvious that this IFU type produces the
highest density of spectra on the detector, making it a good choice for IFUs requiring a
large FoV and many spatial elements.
Extracting the spectra

Keeping in mind the different characteristic appearances of IFU types, it is useful to
address the critical step of data reduction, which is required to produce data ready for
analysis (a comprehensive discussion is presented by Turner in chapter 2 of this vol-
ume). Unlike direct imaging data – from which the observer can generally recover some
fundamental properties – 3D spectroscopy produces raw data that are very difficult to
understand, let alone interpret even only qualitatively. In order to do this, it is first
necessary to extract the spectra from the raw detector frame. Figure 1.14 illustrates
this step with an example from the MPFS instrument (Multi Pupil Field Spectrograph;
Si’lchenko and Afanasiev, 2000). The CCD frame to the left is a continuum flat-field
exposure, showing 16 groups of 16 spectra each, the dispersion direction being horizontal
(left to right). The horizontal black line in the center indicates the spectrum, whose inten-
sity with wavelength is plotted to the upper right. The combined effects of instrumental
transmission, vignetting, CCD QE and the grating blaze function lead to a pronounced
drop of recorded intensity from right (‘red’ wavelengths) to left (‘blue’).
Similarly, the plot to the lower right shows a vertical profile (column) near the center.
As the CCD has 1024 × 1024 pixels; the total of 256 spectra is distributed such that each
spectrum takes up 4 rows of pixels. Since the FWHM of a spectrum perpendicular to
12 Martin M. Roth
Figure 1.14. Example: extraction of MPFS spectra from raw CCD frame.
the direction of dispersion is approximately 2 pixels, and given the more or less Gaussian
profile of the spectrograph point-spread function (PSF), there is considerable overlap
between adjacent spectra. The average inter-spectra minimum intensity is ≈ 65% of the
average peak intensity, while the spectrum-to-spectrum intensity variation is > 20%.
Ideally, the data reduction algorithm must:
(1) find the exact location of all spectra (tracing);
(2) extract the flux per wavelength bin for each spectrum;
(3) compensate for different amount of cross-talk introduced by the overlap of spectra;
(4) apply a ‘fiber-flat-field’ correction to compensate for the variation of throughput
over different fibers; and
(5) rebin all spectra to a common wavelength scale, before the set of spectra can be
used to reconstruct maps, etc.
All or at least part of these steps are actually necessary at the telescope during observing if
the observer wishes to find out whether the IFU was centered correctly on target, whether
the expected signal-to-noise (S/N) has been reached, and whether other indicators of data
quality are met. Note that the data reduction process is a non-trivial exercise, which is
one of the reasons why it took quite a long while for 3D spectroscopy to develop from
the stage of prototypes to the break-through as common user technique.
Data representation, data formats

As already pointed out in Section 1.2.1, the tempting concept of the ‘datacube’ may be
misleading, which is particularly true when considering the representation of 3D data in
digital form. The generic result of data reduction is a set of spectra, associated with a
corresponding set of spaxel positions. The spaxel coordinate system may or may not be
orthonormal, but in the most general case it is not. It is only through the process of inter-
polation in the spatial coordinate system that arbitrary IFU geometries are converted
to orthonormal, i.e. a datacube compatible form (atmospheric refraction not considered;
see Section 1.2.2 below). Interpolation, however, inevitably incurs loss of information.
This is why the Euro3D consortium has introduced and promotes a special data for-
mat for transportation of reduced 3D data which is different from the seemingly simple
application of the standard FITS (flexible image transport system) NAXIS = 3 format,
which has in fact emerged from radio astronomy, where datacubes are a natural data
product (Wells et al., 1981). The Euro3D data format (Kissler-Patig et al., 2004) avoids
1 2 # l ID
o rd ord oup axe
r
Co Co G Sp
ID th x
u m ion X eng nde
tr t A l i
ec lec SP ec art
Sp Se N Sp St
Figure 1.15. The Euro3D data format (from Kissler-Patig et al., 2004).
this latter step of interpolation, merely assuming that the basic steps of data reduction
have been applied to remove the instrumental signature, but otherwise presenting the
data as a set of spectra with corresponding positions on the sky. Figure 1.15 illustrates
the spaxel-oriented approach of the Euro3D (E3D) FITS data format.
Data analysis
3D data analysis comprises the inspection of data quality and the determination of scien-
tifically useful physical quantities. It needs to be clearly distinguished from data reduc-
tion, which should rather be considered instrumental, and as such (almost) part of the
data acquisition process. Some conventional goals of 3D data analysis are: reconstruc-
tion of maps at certain wavelengths, line-ratio maps, velocity fields, co-adding/binning
of spectra, sky/background subtraction, spectrophotometry, etc. These topics will be
addressed extensively later in this volume. At this point it is, however, appropriate to
point out that there is a considerable level of complexity involved if one compares these
with well-established astronomical observing techniques. Although significant progress
has been made over the years, it seems fair to state that 3D spectroscopy as a general
technique is still in its infancy. In an attempt to popularize 3D spectroscopy in the astro-
nomical user community, the Euro3D consortium has made an effort to develop concepts
to facilitate the process of generating science results from 3D data. A key instrument
to this end is a 3D visualization tool, which can be described as a general graphical
user interface to 3D data. The E3D programme (Figure 1.16) as a result of this effort,
developed by S. Sánchez, is extensively described in this book.
1.2.2 Comparison with classical techniques

3D spectroscopy – as opposed to conventional slit spectroscopy – has a number of unique
advantages, which can be summarized as follows, and which will be described in more
! a posteriori advantage, pointing;

detail below:
! absence of slit effects;

14 Martin M. Roth
Figure 1.16. The Euro3D Visualization Tool (Sánchez, 2004; Sánchez et al., 2004a).
a) b)
Figure 1.17. a) ‘A posteriori’ advantage of 3D; b) unlike in slit spectroscopy, it is not necessary
to center the object exactly within the aperture of the instrument.
!
!
atmospheric refraction;
!
spatial binning, low surface brightness;
!
differential spectrophotometry;
!
crowded field 3D spectroscopy;
ultra-deep faint object 3D spectroscopy.
A posteriori advantage, pointing

This advantage is based on the fact that an IFU within its useful FoV has no physical
aperture, so that there is no need to accurately center a given target of interest. The
importance of this property has been highlighted in, for example, papers of Bacon et al.
(1995) on the double nucleus of M31, or Mediavilla and Arribas (1993) on the offset AGN
in NGC 3227, where the physically important feature (peculiar kinematics, emission line
region) were in fact not coinciding with the intensity peak in broad-band light, which on
a standard telescope acquisition camera would normally tend to suggest where to center
the slit.
Slit effects
Slit effects are essentially light losses caused by the finite width of a spectrograph slit.
The severeness of the effect is a function of seeing FWHM in relation to the slit width
(Jacoby and Kaler, 1993), and a function of wavelength. Bacon et al. (1995) have pointed
a) b)
Figure 1.18. Slit effects: a) a wide slit collects a large fraction of target flux; however, it
also collects a large amount of background light. b) A narrow slit minimizes the background
contribution (and also increases spectral resolution), but at the same time reduces the recorded
target flux. No such effect occurs in an IFU.
a) b)
Figure 1.19. a) In slit spectroscopy, the effect of atmospheric refraction can be minimized by
aligning the spectrograph slit with the atmospheric spectrum (parallactic angle). b) Because the
seeing FWHM varies with wavelength, flux calibration is affected by systematic errors; however,
this is not so for IFUs.
out that a surface brightness gradient across the slit in the direction of dispersion causes
a bias in the instrumental profile, which in turn translates into an error on the wavelength
scale. 3D spectrographs based on the micropupil principle are totally unaffected by this
problem, which is why these instruments are superior tools to measure subtle velocity
field variations on small scales, e.g. in the nuclear regions of galaxies.
Atmospheric refraction
Atmospheric refraction is due to dispersion in the atmosphere, which causes a different
angle of deflection as a function of wavelength for rays at oblique angles of incidence, i.e.
for objects at airmass > 1. For point source images, the effect is equivalent to creating
an extended spectrum instead of a well-defined spot in the focal plane of the telescope.
The spectrum is extended in the direction of the parallactic angle, i.e. a great circle
through the zenith and the target, as projected on the sky. Atmospheric refraction is
a function of air temperature, pressure and relative humidity. The effect is described
quantitatively by Filippenko (1982), who advised that a spectrograph should be placed
such that the slit orientation is along the parallactic angle or, in other words, that all
light of the atmospheric spectrum falls into the slit aperture. Since, however, seeing is
a function of wavelength, different fractions of total stellar flux are recorded along the
atmospheric spectrum. Due to this fact, spectrophotometric calibrations by means of a
flux standard are affected by systematic errors whenever the standard is observed at an
airmass that is different from the airmass of the object. None of these problems is an
issue for 3D spectroscopy, at least whenever an IFU with a fill factor close to unity is
employed. Contrary to a slit, the effect of atmospheric refraction translates merely into
16 Martin M. Roth
E 131 4612 Å
b) E 131 7440 Å
Figure 1.20. Maps of standard star exposure at two different wavelengths: a) blue; b) red. The
offset due to atmospheric refraction and the wavelength dependence of the seeing FWHM are
clearly visible. Data obtained with the GMOS-IFU at GEMINI-S (courtesy of J. Turner).
Figure 1.21. Atmospheric refraction in the datacube picture: two maps at two different wave-
lengths are shown as orthonormal grids. Owing to atmospheric refraction, the centroid of a
stellar image experiences a translation from one wavelength to another. Note that, in the most
general case, two exposures obtained at different times (and, therefore, at different airmass and
hour angles) are affected by different translation vectors. Hence, the combination of two datasets
requires a transformation to a common coordinate system, i.e. a correction of the translation
due to refraction.
a shift in x,y as a function of wavelength in the familiar picture of a datacube; however,

no flux is lost due to light falling outside an aperture (Figure 1.20 and Figure 1.21). The
effect on 3D spectroscopy was first described by Arribas et al. (1995) and Emsellem et al.
(1996). Note that it is atmospheric refraction that renders the rigorous definition of a
datacube inappropriate: Figure 1.21 suggests a data rhomboid rather than a datacube.
Because the atmospheric translation vector is never the same for a series of exposures
taken at different times, the combination of datasets for the purpose of improving S/N,
is a non-trivial exercise, involving rebinning and dithering techniques.
Spatial binning
From direct imaging data it is known that poor S/N in regions of low surface brightness
can be improved by binning pixels into superpixels, which contain more signal and,
hence, better S/N. Whenever it is justified to sacrifice high spatial resolution, one can
trade spatial resolution for S/N. The same method can be applied to 3D spectroscopy
by binning spaxels in regions of low surface brightness, where no dramatic variations
are expected to play a role. The gain scales with the square root of the number of
spaxels which are co-added into one superspaxel. This technique has been refined to
what is called ‘adaptive binning’ (Cappellari and Copin, 2003), which is an algorithm
that automatically generates co-added groups of spaxels through the use of a Voronoi
tessellation. The regions are shaped by design such as to share the same level of S/N.
Differential spectrophotometry
As outlined already above (under the headings of Slit effects and Atmospheric refrac-
tion), 3D spectroscopy is, in principle, the ideal method to perform spectrophotometry.
Spectrophotometry, however, is severely compromised under non-photometric observing
conditions. Provided the observer is in the position to find a secondary or tertiary flux
standard near the position of the target, one can apply the technique of ‘differential spec-
trophotometry’, analogous to the development of differential photometry (Barwig et al.,
1987). Using this technique, it is possible to continue observing even through thin clouds
and to compensate for transmission variations through the simultaneous observation of
the standard star. Novel 3D spectrographs with large FoV, or instruments with deploy-
able IFUs, or 3D spectrographs incorporating a direct imager for field acquisition and
guiding are ideally suited to adopt this technique.
Crowded field 3D spectroscopy

Spectroscopy of individual sources embedded in densely populated stellar fields, or in
bright nebular emission, or in a combination of both is a notoriously difficult undertak-
ing, where the precise subtraction of the background component is often a challenge.
Estimating the background at the position of a target is particularly difficult for slit
spectroscopy, where there is only information in one dimension (= along the slit) about
the two-dimensional surface brightness distribution, which is often quite variable in the
spatial domain, as well as with wavelength. Building on the established techniques of
PSF-fitting CCD photometry, it has been shown that the same methods applied to data-
cubes yield superior results when trying to disentangle point sources from a bright back-
ground (Roth et al., 2004b). In practical terms, this is accomplished by processing all
monochromatic ‘slices’ of a datacube one by one. The PSF fitting technique can be even
further enhanced in the presence of high resolution images (e.g. from space), where the
centroids of point sources can be used to iteratively deconvolve the blended features
contained in a datacube (Hook and Lucy, 1993; Becker et al., 2004).
Ultra-deep faint object 3D spectroscopy

Arguably among the most spectacular data products from observational Astronomy are
the Hubble Deep Field images and subsequent similar results, providing the deepest views
into the Universe ever obtained (Williams et al., 1996). These very deep images with
total exposure times of on the order of ≈ 100 hours have only been possible owing to the
linearity and excellent quality of CCD images, which were resampled and stacked using
the sophisticated ‘drizzle’ algorithm (Fruchter and Hook, 2002). To perform similar deep
observations with 3D spectroscopy was first suggested by Roth et al. (2000), and proposed
18 Martin M. Roth
a)
b)
Figure 1.22. Crowded field 3D spectroscopy: the LBV candidate star B416 in M33. a) Hubble
Space Telescope (HST) field, used for deconvolution technique; b) reconstructed map as inferred
from INTEGRAL data. See Fabrika et al. (2005).
for an actual second generation very large telescope (VLT) instrument by Bacon et al.
(2002). It is worth mentioning that the data reduction and stacking process of ground-
based 3D data are significantly more complex than the same process for direct images
from space. Note in particular the complication of different amounts of atmospheric
refraction for datasets obtained at different times (see Figure 1.21). However, with the
validation of concepts for faint object 3D spectroscopy with real observations and stacking
data from many hours of exposure time (e.g. Bower et al., 2004; Roth et al., 2004b), the
development of MUSE (Multi Unit Spectroscopic Explorer, a dedicated faint object 3D
spectrograph for the European Southern Observatory – VLT) is now underway (Bacon
et al., 2004). The deep surveys planned with this instrument are expected to discover
faint Lyα emitters within the 1 arcmin FoV at unprecedented flux levels down to ≈
3 × 10−19 erg s−1 cm−2 .
1.3 A brief history of 3D spectroscopy

Since the early 1980s, the desire to simultaneously acquire spectral and spatial informa-
tion in a single exposure motivated various groups to develop integral field spectrographs
for astronomical telescopes. According to the definition given in Section 1.2.1, the goal
was to obtain three-dimensional datasets (datacubes), which should be generated from
a single exposure under identical observing conditions, contrary to datacubes that are
assembled through a scanning process involving multiple exposures. As an example for
the latter method, Wilkinson et al. (1986) investigated the kinematics in the Cen A galaxy
(NGC 5128) with a datacube generated from a series of offset long-slit exposures. They
used the RGO Spectrograph at the 3.9 m Anglo-Australian Telescope (AAT) with a
V1200 grating, a slit width of 0.3 mm, and the Image Photo Counting System (IPCS) as
detector. Along the slit, the sampling was 2.2′′ per spatial element over a total length of
61 elements. With a total exposure time of 13,327 s for the entire datacube, 71 telescope
offset positions in increments of 2.4′′ were needed. The spectral domain was sampled with
a total of 1650 increments of 0.5 Å. Resulting from these observations, the velocity field
over the face of the galaxy was derived and conclusions drawn concerning the kinemat-
ics of the stellar component with regard to the warped dust lane in this nearest large
elliptical galaxy, which has just experienced a major merger event. The example demon-
strated vividly the need for spatially resolved area spectroscopy, which indeed could
be fulfilled with the long-slit scanning technique. There were other successful applica-
tions for different science problems – e.g. the study of planetary nebulae (Balick et al.,
1987) – but it was soon realized that the scanning techniques have their shortcomings. Not
only was the sequential data-taking process cumbersome and inefficient, but it was also
sensitive to variable atmospheric conditions. For example, an identical PSF throughout
the FoV was found to be important for measurements of the two-dimensional kinematics
in the nuclei of galaxies, requiring high spatial resolution, as pointed out by Bacon et al.
(1995). Therefore, one would ideally like to obtain the entire datacube from a single
exposure.
However, before the time was ripe for true integral field spectroscopy, another scan-
ning method was introduced which did indeed provide a stable PSF for any map derived
from the datacube (albeit not necessarily identical from wavelength to wavelength step),
namely the imaging Fabry–Perot (FP) interferometer. In fact, applications of this tech-
nique go back to the early 1970s, e.g. when Tully (1974a–c) used the Maryland pho-
tographic Fabry–Perot Interferometer at the KPNO (Kitt Peak National Observatory)
2.1 m Telescope to determine the Hα velocity field in M51. For this work, photographic
plates with IIa-O emulsion, and a two-stage Carnegie image intensifier with an S20 pho-
tocathode were employed, requiring typical exposure times of 0.25−2 hours plate−1 (10 h
total). The remarkable result was the kinematic mapping of the galaxy over a FoV as large
as 12 arcmin in diameter with a spatial resolution of 2′′ , and a spectral resolution of 0.5 Å,
corresponding to a velocity resolution of 27 km s−1 , over a free spectral range of 4.2 Å. The
sensitivity limit of these observations was 10−15 erg s−1 cm−2 arcsec−2 , and radial veloci-
ties could be measured wherever the Hα intensity exceeded 10−14 erg s−1 cm−2 arcsec−2 .
Tully was able to fit a three-component model of the mass distribution, comprising disk,
bulge and the nucleus. Subsequently, FP tunable filter instruments became more and
more popular, e.g. the famous TAURUS and TAURUS-2 at WHT and AAT, respec-
tively (Atherton et al., 1981, 1982; Bland-Hawthorn and Taylor, 1994), HIFI at CFHT
(Canada–France–Hawaii Telescope), and the prime focus FP at the Calar Alto 3.5 m
(Meisenheimer and Hippelein, 1992). The major advantage of these instruments was a
large FoV, high spectral resolution and large throughput. Several disadvantages – such
as ghost images, stray light, interference artifacts, problems with flux calibration, the
narrow useful wavelength range for a given etalon, and a relatively complicated data
reduction process – have prevented these systems from becoming standard instruments
for the common observer. Also, FP imagers are not true 3D spectrographs in the sense
that they require scanning.
By the time that FP instruments had their technological breakthrough and the con-
cept of IFS was more commonly appreciated, optical fibers had been discovered as a
means of increasing the multiplex of spectroscopy towards making large spectroscopic
surveys feasible: The first astronomical spectrum through a fiber was taken in 1978, and
the first fiber-optical multi-object spectroscopy was performed using the Steward Obser-
vatory 2.3 m telescope just one year later (Hill et al., 1980; see also a review by Hill,
1988). Almost at the same time Vanderriest (1980) used the design flexibility offered
by optical waveguides to build the first fiber bundle prototype, which was installed at
the Mauna Kea 2.2 m telescope. It comprised a total of 169 fibers, covering a hexagonal
footprint of 27′′ × 43′′ with a fill factor of 75% and a projected fiber diameter of 0.42′′
on the sky. Thirty-six additional fibers, arranged in four blocks away from the central
FoV, were foreseen for sky background subtraction. While this setup was coupled to an
existing, however ill-adapted, photographic image-tube spectrograph, in its early form it
was already incorporating all major features that were later on included in subsequent
dedicated fiber bundle IFS. Because of the technical constraints of the experimental
prototype at the time, the scientific outcome (3D spectra of 3C 120) was somewhat
limited, but it is interesting to note that about a quarter of a century later the use
20 Martin M. Roth
Figure 1.23. DensePak-2 at the KPNO 4 m telescope.
of a similar fiber bundle, coupled to a dedicated fiber spectrograph with a modern CCD
detector (INTEGRAL at WHT; Arribas et al., 1998a; 1998b) yielded high-quality data
of 3C 120, which allowed the application of image-processing techniques to accurately
disentangle the radio jet emission from the underlying galaxy (Sánchez et al., 2004b).
Another experiment was the DensePak-1 IFU for the KPNO 4 m telescope (Barden
and Scott, 1986), which also had the shortcomings of an early prototype. However, the
early experiments stimulated a whole series of new fiber bundle IFU developments at
different telescopes; for example, DensePak-2 (Barden and Wade, 1988) was a much
improved follow-up at the KPNO 4 m, featuring a rectangular fiber bundle with a total
of 49 fibers over a FoV of 16′′ × 19′′ and a projected fiber diameter of 2.1′′ on the sky;
SILFID (Spectrographe Integral Linarisation par Fibres de l’Image Directe) for the CFHT
(Vanderriest and Lemmonier, 1988); HEXAFLEX/WHT (Arribas et al., 1991), 2D-FIS
(2-Dimensional Fiber ISIS System) WHT (Garcı́a Marı́n et al., 1994); HEXAFLEX-II/
NOT (Garcı́a Marı́n et al., 1994); ARGUS (Vanderriest, 1995). Although fiber bundle
IFUs suffer from incomplete sampling and sometimes irregular arrangement of the fibers,
they have an intrinsically simple structure and can be fitted as add-on modules to exist-
ing bench-mounted spectrographs, which probably explains why they represent the most
common type of IFS in the early days of 3D spectroscopy. A more recent interesting appli-
cation is the SparsePak IFU at WIYN (Wisconsin–Indiana–Yale and National Optical
Astronomy Observatory; Bershady et al., 2004).
It is worth mentioning that when the topic of IFS had gained considerable momentum
in the 1990s, a dedicated International Astronomical Union (IAU) Colloquium was orga-
nized in 1994 (Compte and Marcelin, 1995), after which, superseding the initial phase
of prototypes, a first generation of common user 3D spectrographs emerged at 4 m-class
telescopes.
One of the prominent examples is INTEGRAL at the WHT (Arribas et al., 1998a;
1998b). Unlike several of the early prototypes, INTEGRAL is an optimized fiber bun-
dle IFU, attached to the WYFOS (Wide-Field Optic Spectrograph) fiber spectrograph
at WHT (Bingham et al., 1994). INTEGRAL features three rectangular fiber bundles
at different magnifications for the purpose of accommodating different observing condi-
tions in terms of seeing, and different user requirements (spatial resolution versus FoV).
Each bundle is also furnished with a set of separate fibers for sky background subtrac-
tion. The three bundles have parameters as follows (FoV/projected fiber diameter): SB1
7.8′′ × 6.4′′ /0.45′′ ; SB2 16.0′′ × 12.3′′ /0.9′′ ; SB1 33.6′′ × 29.4′′ /2.7′′ . Besides the different
Figure 1.24. INTEGRAL at the Nasmyth focus of the WHT.
SB1 SB2 SB3 Sk fibers

Sk fibers Sk fibers
f20 mm f20 mm f20 mm

(90.15' ) (90.15' ) (90.15' )
bserving fibers bserving fibers bserving fibers

As vie ed b As vie ed b As vie ed b
the measurer the measurer the measurer
Broken ( 78)
1.485 mm
Broken ( 167)
2.718 mm
(6.69' )
6.731 mm
Broken ( 30)
(12.25' )
(30.34' )
Broken ( 17)
600 µm
200 µm
100 µm (2.70' )
(0.90' )
1.742 mm (0.45' )
(7.95' ) 3.643 mm
(16.42' ) 7.636 mm
(34.42' )
Figure 1.25. INTEGRAL fiber bundles.
choice of IFU magnification, INTEGRAL offers a variety of grating options with dif-
ferent dispersion, wavelength coverage and blaze thanks to the many modes available
in WYFOS (Wide-Field Optic Spectrograph). This property makes INTEGRAL a very
useful common user instrument as it allows observers to a large extent to optimize the
instrumental setup and accommodate to their specific scientific problems.
Fiber bundle IFUs have the advantage that they are conceptually simple and that they
can be attached a posteriori to existing (preferably fiber) spectrographs. It is therefore
probably no surprise that the very first IFUs were indeed based on this principle. However,
a few shortcomings soon led to alternate designs, in particular the inevitable drawback
of fiber bundles having fill factors of less than unity. The incomplete sampling does not
only influence the spatial resolution of seeing-limited observations, but also gives rise to
photometric inaccuracies when observing spectrophotometric flux standards: whenever
an appreciable fraction of standard star flux falls into the gap space between adjacent
fibers, the amount of lost light cannot be straightforwardly accounted for. A way to
overcome this shortcoming was to use an optical device that had already been proposed
by Courtès (1960), namely a lens array in the optical plane of the telescope. While
22 Martin M. Roth
the principle of operation is explained in more detail in Section 1.4.2, suffice it to say
for the purpose of this overview that a contiguous array of microlenses of square or
hexagonal shape does not suffer dead space between adjacent spatial elements, unlike
a conventional fiber bundle. Such a device was therefore thought to be ideal for the
highest spatial resolution of IFS. Using the principle of micropupil imagery and suitable
dispersion in combination with a waveband selecting filter, it could also accommodate a
large density of spectra on the detector, i.e. a larger number of spatial elements. Despite
these advantages, it was not until 1987 when the first lens array IFU ‘TIGER’ saw first
light at the CFHT (Courtès et al., 1988; Bacon et al., 1995). Other implementations
were shortly thereafter developed for the Selentchuk 6-m telescope (Afanasiev et al.,
1990) and in Kyoto (Ohtani et al., 1994). The TIGER instrument was not only ground-
breaking in terms of technological innovation, but also quite successful with scientific
achievements, e.g. the kinematic study of the double nucleus in M31 (Bacon et al., 1994),
to name just one example. A subsequent follow-up development built by the same group
in Lyon, France, was the OASIS (Optically Adaptive System for Imaging Spectroscopy)
IFS, which was optimized for very high spatial resolution in combination with the PUEO
(Probing the Universe with Enhanced Optics) adaptive optics system at CFHT. Although
adaptive optics (AO) operation was difficult to accomplish with natural guide stars only,
this development has pointed to what is now generally thought to have become a leading
track towards novel spectroscopic instrumentation at large and extremely large telescopes
(ELT), i.e. AO-assisted integral field spectroscopy (see below). It has even been claimed
that modern adaptive optics spectroscopy in general only makes sense when performed
with an IFU.
Although not strictly speaking a common user instrument, a further development of
the TIGER principle has led to what became extremely successful in terms of science
output, namely the SAURON IFS at the WHT (Bacon et al., 2001). This instrument
was very specifically designed and is operated for a particular science case, i.e. absorp-
tion or emission line kinematics and stellar population analysis in elliptical galaxies. An
overall simple opto-mechanical layout with practically no moving parts has made the
instrument very stable and reliable for what is essentially a single-purpose mission; it
has accomplished this with superb results (e.g. de Zeeuw et al., 2002). In the trend of
an increasing number of 3D spectrographs becoming available at 4 m class telescopes,
SAURON may well be counted in the category of post-prototype instrumentation, albeit
having the status of a private instrument.
In the process of the development of the baseline designs for IFS as discussed above –
i.e. fiber bundle versus lens array – it became clear that the two are largely complementary
to each other: large wavelength coverage, relatively coarse spatial sampling and moderate
number of spectra, versus moderate wavelength coverage, optimal spatial sampling and
large number of spectra. However, it was found that for certain observational problems
it would be desirable to combine the superior spatial sampling properties of lens arrays
with the large wavelength coverage of fiber bundle IFS. As a result, hybrid ‘lens array –
fiber bundle’ IFUs were developed (see Section 1.4.3). The first implementation of this
type of IFU happened in the MPFS (Multi-Pupil Field Spectrograph) instrument at the
prime focus of the Selentchuk 6 m telescope in 1997. Other IFSs using this principle are
SPIRAL/AAT (Kenworthy et al., 2001), PMAS/Calar Alto (Roth et al., 2005; Kelz et al.,
2006) and several other instruments at 8 m-class telescopes as discussed below.
Fiber and lens-array-based designs have governed the layout of the first generation
of IFSs. Given the rather limited number of spaxels in comparison with direct imaging
detectors, and also considering the inefficient use of detector space, the slicer principle
was introduced with the ‘3D’ instrument (Weitzel et al., 1996). This IFS was designed
Figure 1.26. SAURON at the WHT, La Palma.
Figure 1.27. PMAS at the Calar Alto 3.5 m telescope features two IFU options: a fibercoupled
lens array IFU, and a bare fiber bundle IFU.
24 Martin M. Roth
a) b)
Figure 1.28. a) Compact GMOS-IFU. b) VIMOS-IFU, the largest IFU to date, providing a
total of 6400 spectra.
for the J, H and K bands in the near infrared (NIR). Since the available detector size
was small (256 × 256 pixels), it was felt that the number of fiber or lens array spectra
that could be projected onto the available chip would be too small to warrant the effort
of building a conventional IFU. As an alternative, the slicer method generates a stack of
long-slit spectra, corresponding to a series of parallel slices in the image plane. Contrary
to the fiber and lens array IFUs, the sky is thus imaged directly onto the detector, whereas
in the former case the fiber end faces or micropupils are projected onto the focal plane.
The detector area is therefore optimally filled with contiguous spectra and no space is
lost due to gaps. The optomechanical design of slicers is less critical at long wavelengths,
but more difficult in the optical or UV as far as performance in terms of image quality,
edge effects and stray light is concerned. A few more NIR slicer instruments were built
for 4 m-class telescopes, but none for operation in the optical (PIFS/Hale 5 m, Murphy
et al., 1999; UIST/UKIRT, Ramsay Howatt et al., 2004).
The history of IFS as discussed so far included the very first prototypes and a first gen-
eration of facility instruments for 4 m class telescopes. The latter already represented a
significant step towards establishing 3D spectroscopy as a common user technique. In the
late 1990s the light collecting power, superb image quality and demand for new instru-
mentation of new 8–10 m class telescopes prompted yet another generation of IFSs, some
of which were designed as IFUs in the original sense, i.e. add-on modules in front of con-
ventional slit spectrographs (e.g. GMOS-IFU: Gemini Multi-Object Spectrograph – IFU;
Allington-Smith et al., 2002; VIMOS-IFU/VLT: Visible Imaging Multi-Object Spectro-
graph – IFU/VLT; Le Fevre et al., 2003), while others are dedicated 3D spectrographs
(SINFONI/VLT, Eisenhauer et al., 2003; OSIRIS/KECK, Larkin et al., 2006). These
instruments usually come with their own data reduction pipeline, which is an important
prerequisite for opening this non-trivial observing technique to the common user.
Future facilities are presently planned for ground-based and for space telescopes. The
most prominent new developments for the ground are MUSE as a second-generation
VLT instrument, and VIRUS (Visible Integral-Field Replicable Unit Spectrograph) for
HET (Hobby–Eberly Telescope). These new facilities are planned to accommodate a
total of as many as 90,000 and 30,000 spectra, respectively. As a consequence, they are
quite large and highly complex, requiring novel approaches for manufacture and integra-
tion, e.g. replicable devices which can be built in small industrial series, thus reducing
risks and saving costs. In space, JWST is planned to host – for the first time – IFUs
above the atmosphere, namely as part of the NIRSpec (Near-Infrared Spectrograph)
and Mid-Infrared Instrument (MIRI), covering the 0.6−5 µm and 25−27 µm wavelength
ranges respectively (Arribas et al., 2005). Of those future instruments, MUSE will come
Figure 1.29. MUSE at the Nasmyth platform of the VLT (CAD layout).
reasonably close to the ideal of a conventional imager, however creating a full one-octave
spectrum in each pixel. Over the wavelength range of 465−930 nm, MUSE will accommo-
date a square FoV of 1 arcmin with a spatial sampling of 0.2 × 0.2 arcsec2 in its wide-field
mode. The narrow-field mode will cover 7.5 × 7.5 arcsec2 with a sampling of 0.025 arcsec2 .
The instrument will be assisted by a ground-layer correcting AO system with 4 laser
guider stars and an additional natural guide star (Bacon et al., 2004).
1.4 Technical approaches

In the following, we will discuss in more detail several relevant aspects of technologies
! optical fibers;
that are in use for 3D instrumentation; namely:
! lens arrays;
! lens arrays coupled to fiber bundle;
! slicers;
! special techniques;
! non-IFS 3D instruments;
! 3D detectors; and
! figures-of-merit.
1.4.1 Optical fibers

Historically, optical fiber bundles represented the first and most simple technology to
build up IFUs, often as front ends to existing slit spectrographs. Consequentially, some
of the early prototypes were not optimized in terms of efficiency, stability or other param-
eters. However, a whole family of specifically designed facility instruments following the
first generation of prototypes has convincingly demonstrated that optical fibers are quite
useful components for 3D spectroscopy. Standard multimode optical fibers are thin waveg-
! the fiber core made of high refractive index glass;

uides basically consisting of three components:
! the cladding as a low refractive index skin around the core; and
! a protective buffer normally made of some sort of plastic material, e.g. polyamide.
26 Martin M. Roth
Figure 1.30. Fiber 1.
Figure 1.31. Fiber bundle.
Figure 1.32. Focal ratio degradation.
They act as wave guides owing to the effect of total internal reflection, meaning that rays
at grazing incidence onto the core–buffer interface will be reflected back into the core.
Multiple reflections inside the core lead to a scrambling of the input light pattern. Rays
incident onto the fiber input with inclination angles larger than that given by a certain
critical angle (numerical aperture, NA) will no longer undergo total internal reflection,
but will rather be absorbed in the buffer. Figure 1.31 shows a microscopic view of the
polished front face of a bundle formed from several such fibers which were bonded together
to make up a hexagonal densest package. In this picture, the bright circular areas are the
fiber cores, surrounded by a dark ring (the cladding), and another material (the buffer).
Note that this fiber bundle IFU has an incomplete fill factor, as only a fraction of the
area is acting as a light-collecting aperture. One of the most important disadvantages
of fibers is their property of not normally retaining the input f-number of the incoming
beam (Figure 1.32). The emerging light cone has a larger opening angle than the input
beam. The effect is called focal ratio degradation (FRD). It is very undesirable as it leads
Figure 1.33. Several views of the fiber-optical IFU ‘PPak’ of PMAS (Kelz et al., 2006).
to overfilling the spectrograph collimator and light losses. Moreover, since the amount
of FRD depends on external parameters like mechanical or thermally induced stress, the
output f-number is not constant with time unless the fiber duct from the IFU to the
spectrograph is extremely stable. In certain cases, variable FRD is capable of creating
noise, which limits the ultimate S/N obtainable, even under conditions with a lot of
light and good photon statistics (Parry and Carrasco, 1990; Baudrand and Walker, 2001;
Schmoll et al., 2003).
In summary, fiber-optical IFUs have the following properties: Fibers allow for a very
flexible installation of an IFU, and they accommodate bench-mounted spectrographs,
in particular, which is a strong argument for stability. There are a few problematic
fiber properties, which – with proper care and a suitable design – can be handled fairly
well. Fiber-optical IFUs have typically relatively few spaxels, but – depending on the
spectrograph used – a large number of spectral bins and a wide wavelength coverage.
Fiber bundle IFU can be built with large spaxels to cover a sizeable area on the sky, which
is, for example, extremely beneficial for low surface brightness spectroscopy. One of the
more serious drawbacks for certain applications, especially when high spatial resolution is
intended, is the non-contiguous sampling due to inevitable gaps in the bundle. However,
contrary to common belief, fiber-optical IFUs can be built with high efficiency.
1.4.2 Lens arrays

The suggestion of using lens arrays for astronomical observations goes back to Courtés
(1960); however, it took quite a while until the first implementation appeared at the
telescope (Bacon et al., 1995). Figure 1.34 shows an example of a lens array that has
a hexagonal geometry (OASIS-IFU). The first 3D spectrograph of the lens-array-type
was TIGER, the famous prototype for this class of instrumentation. At first glance, the
lens array has one significant advantage over fiber bundles, namely the absence of large
gaps between the spaxels. Although even for lens arrays there are small dead zones at
the interfaces between adjacent lenslets, the affected area in proportion to the total lens
28 Martin M. Roth
Figure 1.34. Lens Array of the OASIS 3D Spectrograph.
Focal plane Fabry lens
Figure 1.35. Principle of a micropupil (not to scale).
Magnified focal plane of telescope

Field lens
Focal plane of telescope
Pupil image
Magnifying lens
Macrolens array
Figure 1.36. Optical layout of a lens array with fore-optics (adapted from Dean, 2002).
area is negligible for most lens arrays. Lens arrays are therefore superior devices for high
spatial resolution studies of objects on small scales. Another significant advantage of lens
array IFUs is often overlooked, namely the principle of micropupil imagery. Instead of
imaging the dispersed spaxels of the IFU, it is the microscopic entrance pupil behind each
lenslet (= spaxel) of a lens array that is imaged onto the detector (hence the term ‘micro-
pupil’). The practical advantages of micropupils for precision spectroscopy are discussed
in Section 1.2.1.
The sketch in Figure 1.35 illustrates the optical principle, which is well known from
the technology of one-channel photoelectric photometry, where a so-called ‘Fabrylens’
images the entrance pupil of the telescope onto the photocathode of a multiplier. Instead
of having to accept the noise caused by seeing-induced image motion of a stellar image
on the photocathode, the static pupil image is unaffected by atmospheric effects or guid-
ing errors. Translating this macroscopic arrangement to the scale of a lens array in
Figure 1.37. Example for lens array spectra from SAURON.
combination with some fore-optics having the proper focal length, the result is a set
of microscopic pupils behind the lens array, with one micropupil for each lenslet. The
size of the micropupils is designed to match the optimal slit width of an imaging spec-
trograph with a sufficiently large FoV to encompass the physical size of the lens array.
The arrangement is then very similar to a multi-object slit spectrograph of the EFOSC
(ESO Faint Object Spectrograph and Camera) type (Buzzoni et al., 1984), except that
no punch plate is necessary as the slits are defined by the micropupils, and the micro-
pupils are aligned on a totally regular grid. In order to avoid the overlap of spectra com-
ing from neighboring micropupils, the spectral coverage per spectrum must be truncated
to approximately the pitch of the lens array, which can be done by using a narrowband
filter in the fore-optics train. Moreover, in order to make more efficient use of detector
area and to maximize the spectral coverage, the dispersive element (grating, grism) is
slightly rotated, resulting in a diagonal arrangement of spectra in the spectrograph focal
plane. Figure 1.37 shows a real example from the SAURON instrument (Bacon et al.,
2001) as magnified views from:
a) the undispersed micropupils;
b) a continuum flat-field exposure;
c) a spectral line lamp calibration exposure; and
d) an astronomical object exposure.
In summary, lens array IFUs can, in principle, accommodate a large number of spax-
els at the expense of wavelength coverage, and some possible problems with truncated
spectra near the edge of the detector. The fact that there is contiguous sampling and no
gaps makes them very suitable for high spatial resolution observations. Thanks to the
micropupil principle, the telescope focal plane is decoupled from the detector, which is
a significant advantage for precision spectroscopy. The compact geometry of the spectra
allows for compact optomechanical designs and also for highly efficient optical systems.
A certain drawback of lens array IFUs is the presence of diffraction causing crosstalk
between neighboring spectra, which occurs at different wavelengths (Lee et al., 2001;
30 Martin M. Roth
Figure 1.38. CIRPASS fiber bundle.
Roth et al., 2005). This undesirable effect can, however, largely be corrected for by soft-
ware in the process of data reduction.
1.4.3 Lens array – fiber hybrids

There are applications requiring the sampling properties of a lens array; however, with a
wavelength coverage that is more extended than can be accommodated by the truncated
spectra of a classical TIGER-type IFU. A solution for this requirement has been realized
by combining a lens array with a fiber bundle, where each optical fiber is centered on
one micropupil at the exit plane of the lens array. The versatility of the fibers can then
be conveniently used to reformat the lens array geometry to a pseudo-slit, thus using the
entire width of the detector in the direction of dispersion. Figure 1.38 shows a picture
from the fiber bundle attached to the back of the macro-lens-array of the CIRPASS
(Cambridge Infrared Panoramic Survey Spectrograph) instrument (Parry, 2004).
As opposed to the pure lens array or pure fiber bundle IFU, the hybrid solution is
affected by additional light losses that depend very much on the image quality of the
lens array, the precision of the alignment of fibers with regard to the micropupil position,
scattered light, diffraction, focal ration degradation, etc. There are instruments that
feature an IFU as an add-on module to a classical slit (or MOS) spectrograph, in which
case often another set of microlenses is applied to the fiber output ends in order to match
the fiber output to the spectrograph beam. In such cases, there are more light losses
associated with the additional optical elements. On the positive side, the finite apertures
of the fibers in the micropupil plane operate like a spatial filter, blocking diffraction
spikes and scattered light coming from the lens array, which, in the case of TIGER type
of instruments, would be propagating into the spectrograph. A detailed description of the
relevant effects for the example of the PMAS instrument is given by Roth et al. (2005).
To summarize, lens-array-fiber hybrids are good for flexible installations; they can
accommodate bench-mounted spectrographs, but suffer from problematic fiber proper-
ties (which can be handled with proper care). These IFUs typically have relatively few
spaxels, but a large number of spectral bins and wide wavelength coverage. They provide
contiguous sampling and have no gaps between spaxels, making them good devices for
observations that require high spatial resolution.
Inpu t from the

telescope
T elescope
focal plane
Output to
spectro meter
Figure 1.39. Slicer principle of operation (from Weitzel et al., 1996).
1.4.4 Slicers
The IFU types discussed in Sections 1.4.1 to 1.4.3 have one feature in common: they
do not use detector space very efficiently. This obviously has to do with the fact that
re-imaging of fiber end faces, or micropupils, respectively, which incurs a finite image
quality, and, in fact, a broadened image of those ‘slit elements’, makes it necessary to
provide for a minimal spacing of spectra, unless the crosstalk from the broadened profiles
is not regarded as critical. If this is indeed the case, one spectrum would typically cover
four rows of spectra in the spatial direction, assuming that Nyquist sampling was given
by design. Such a configuration consumes a factor of two more CCD pixels than required
in the optimal case, where one spaxel is critically sampled, i.e. it uses just two rows of
CCD pixels. If maximization of the number of spaxels is a primary design goal, then the
slicer concept is significantly more efficient than the competing IFU types. Slicer IFUs
are conceptually based on the optical elements designed to circumvent slit losses in con-
ventional spectrographs; e.g. the Bowen–Walraven image slicer. The image slicer for 3D
spectroscopy was first introduced for the MPE (Max Planck Institut für extraterrestrische
Physik) instrument ‘3D’ (Weitzel et al., 1996), which was designed as a prototype and a
travelling instrument; it was adapted and operated at a variety of 4 m-class telescopes.
3D operates in the NIR, which precludes the use of fibers because of stability issues in a
cryogenic environment. Conversely, the slicer as a relatively small optical element is well
adapted for operation in a cryostat at low temperatures and allows for an overall compact
optomechanical design. The principle of operation of a slicer is shown in Figure 1.39. Light
coming from the telescope is deflected by a stack of tilted mirrors (the ‘slicer’ proper),
which are very narrow in one direction and quite extended in another direction. These
thin mirrors ultimately act as slits, or, in other words, as spaxels, dissecting the focal
plane into narrow and long strips. Another set of mirrors accepts the deflected beams
corresponding to each slice and projects the outgoing beams such that an (almost) linear
arrangement of the slices is created at the entrance of the spectrograph. The benefit of
such a configuration is that the telescope focal plane is directly imaged onto the detector
so that, with the choice of a proper scale, one detector pixel can be interpreted as one
spaxel on the sky. Of all IFU types, this scheme yields the highest packing density on the
detector. However, unlike with micropupil or fiber-based 3D spectrographs, slicer IFUs
do not provide image scrambling and therefore do not share the benefits as discussed
in Section 1.2.1. The first slicers were used for instruments operating in the NIR, e.g.
32 Martin M. Roth
Figure 1.40. Discrete Zerodur slicer element (MUSE prototype).
Reflective coating
r
B
A α
α
π phase
change
α
d
Figure 1.41. FP-principle.
3D, SINFONI. With recent advances in diamond-turning technologies, the manufacture

of metal optics slicers showing surface roughness values < 5 nm has become possible;
this means that metal slicers are also useable in the optical (e.g. MUSE, Bacon et al.,
2004).
In summary, slicers provide high density of spectra, high efficiency, and a compact
optomechanical design for high stability. Slicers have been manufactured from Zerodur,
and from metal as monolithic devices. Slicers are routinely operated in the NIR, and there
are ongoing research and development efforts to optimize metallic slicers for operation in
the visual. In contrast to fiber and lens array IFUs, no scrambling is involved, and the
telescope focal plane is projected directly onto the detector.
1.4.5 Non-IFS 3D instruments

Without going into much detail, some techniques shall be mentioned that are not exactly
3D in accord with our definition; they do, however, produce data with the same char-
acteristics (‘datacubes’) as integral field spectroscopy. These non-IFS 3D instruments
are generally scanning instruments, i.e. the principal difference is that the measurement
process is non-simultaneous.
Fabry–Perot
A very prominent technique, starting from the early days of IFS developments, is the
Fabry–Perot interferometer (FP). The basic principle of interference in a FP is shown in
the sketch of Figure 1.41: a cavity enclosed by a pair of high-quality plane-parallel glass
plates with a reflective coating of reflectance r is acting as a resonator, where rays of
Photomultiplier
tube
Mode selector
wheel
Calibration source Guider eye piece
Flip in mirror
1/8 Beam
Interference filter F1
Focal plane (80 mm diameter) IRIS

Field lens
Interference filter F2 Pupil
Flip in mirror
Detector
Collimator 1/7 B7 393 mm focal length Fabry–Perot Camera lens – 1/2 102 mm focal length
Figure 1.42. TAURUS.
equal angles of incidence experience the same phase superposition, and – depending on
the optical path length difference – constructive or destructive interference. The optical
path length between adjacent rays is given by: 2AB − CD = 2d cos α. The condition for
constructive interference is then:
2d cos α = mλ (1.2)
For every wavelength λ satisfying this condition, the interferometer is transmissive, and
other wavelengths are extinct. Because of the periodicity of the cos-function there are
very many orders m with constructive interference when the spacing d is on the order
of mm. By varying d, one can tune the interferometer to different wavelengths, and
scanning over many values of d is effectively creating a spectrum. Note that in order to
avoid ambiguity owing to the many orders m, it is necessary to use a narrowband filter
to select the wavelength of interest. The width of the passband depends on the etalon
reflectance r, and it can be shown that the spectral resolving power is given by:
√
λ r
= mπ (1.3)
∆λ (1 − r)
A prominent example for a scanning FP imager is the TAURUS instrument (Atherton
et al., 1981; 1982), the layout of which is shown in Figure 1.42. Some of the instrumental
characteristics of TAURUS were as follows. Mode of operation: seeing-limited direct
imaging, FoV = 9 arcmin, spectra resolution: 200 < R < 100,000, IPCS photon counting
detector, wavelength range: 400−700 nm. General features of a scanning FP are: a large
field-of-view, high spatial resolution, non-simultaneous measurements, small wavelength
coverage (generally one spectral line), high spectral resolution.
Fourier transform spectrometer

A special variant of a scanning 3D instrument is the Fourier Transform Spectrometer
(FTS), which is based on the principle of a Michelson interferometer. As opposed to the
scanning FP, which produces direct images, the FTS creates an interference pattern that
as such is incomprehensible to the human eye. It is only through a Fourier transform
performed on a computer that an image is recovered. This instrument type is perhaps
34 Martin M. Roth
Stationary mirror
Beam splitter
Path B
Collimator
Moveable
mirror
Path A
Camera
Detector
Figure 1.43. FTS.
Magnetic field
Base lead Top lead
Top metal film
Substrate
AIOx barrier
Bottom metal film

Illumination
Figure 1.44. Superconductive tunnel junction detector.
less straightforward than other 3D spectrographs, but it has a certain number of advan-
tages. These are quoted with several instrumental parameters of the CFHT FTS as an
example (Bohlender, 1994): the CFHT FTS is a scanning NIR instrument (1−5 µm),
having a high throughput of up to 25%, independent of wavelength. High S/N observa-
tions are supported based on a dual channel beam switching mode, which allows for an
accurate internal sky subtraction before any signal is detected. The instrumental profile
is extremely well controlled (and measurable). The wavelength calibration accuracy is
≈ 100 m s−1 , and the spectral resolution up to 500 m s−1 .
1.4.6 3D detectors
The dream of observational astronomers would be a photon counting detector that sam-
ples all parameters of an incident photon simultaneously; i.e. wavelength, phase and
amplitude, polarization and direction. 3D spectrographs as described above are a hum-
ble yet very powerful approximation towards this goal. However, a compact detector
having these properties would seem to be more appealing intrinsically than the rather
complex layout of the various IFU types. Such detectors do exist, and a prominent pro-
totype development is the Superconduction Tunnel Junction (STJ). This device consists
of a pair of films that are separated by an insulating layer. The films are made of super-
conducting materials like niobium, tantalum, hafnium, etc. The detector is operated at a
temperature below Tcrit ≈ 1K. Owing to the Josephson effect, there is a current through
the junction in the presence of a bias voltage and a magnetic field. The current is sup-
pressed by incident light, where an intrinsic energy resolution is provided due to the
release of a charge from the incident photon. The typical resolving power that can be
achieved, however, is R = 10−100, i.e. very low. This limitation is given by the band-gap
values attained by selected materials, for which research and development efforts are
attempting to provide further improvement. For fundamental reasons, however, spectro-
scopically interesting resolutions one order of magnitude larger than the present values are
not going to become real. Therefore, the STJ will remain a niche application that cannot
replace IFUs. Nevertheless, 3D detectors with spatial resolution have been demonstrated
as arrays of 10 × 10 Josephson cells, forming a real 3D detector. First astronomical obser-
vation with the STJ were presented by Perryman et al. (1999). Their measurements of the
Crab pulsar validated that the STJ has advantages for high time resolution multiband
imaging, especially in the UV. Other developments include so-called Kinetic Inductance
Detectors (KID), which are being developed as large devices with 100 × 100 elements and
R = 30, or Transition Edge Detectors (TES) with 8 × 8, R = 20.
In conclusion, intrinsic energy resolution detectors are still at an early stage of devel-
opment. Their cryogenics to operate at superconduction temperatures around 1 K are
complex and expensive. The devices can only be made to achieve low spectral resolution.
Although they are sensitive in the blue, they are less suitable for the visual and useless
in the NIR. A positive feature is their high time resolution.
1.4.7 Special techniques

Multi-slit masks
A rather simple special slit configuration in front of an imaging spectrograph is capable
of providing pseudo integral field spectroscopy; pseudo in the sense of very low fill factor.
The basic idea is to make a slit mask that has a large number of parallel long slits, each of
which is allowed to generate a spectrum over a limited wavelength interval, which is deter-
mined by the distance to its next neighbors in order to avoid overlap of adjacent spectra.
The wavelength interval can be selected by means of a narrowband filter anywhere in the
spectrograph. The example shown in Figure 1.45 was implemented in the FOCAS (Faint
Object Camera and Spectrograph) instrument at the SUBARU 8.2 m telescope for the
purpose of detecting intracluster planetary nebulae in the Coma Cluster (Gerhard et al.,
2005). The chosen slit configuration yielded 3600 spaxels per field diameter at a spatial
sampling of 0.1 arcsec pixel−1 . The dispersion using the 300B grating was 1.5 Å pixel−1 ,
resulting in 43 spectral bins per spectrum with a pre-filter of 60 Å FWHM, centered
at 5121 Å (redshifted [O III] 5007). The total effective survey area was 12,215 arcsec2 ,
which is 12% of the FOCAS FoV. The remarkable result of a survey in Coma with a
total exposure time of no longer than 3 h and a seeing of 0.6−0.8 arcsec FWHM has
been that planetary nebula candidates with a limiting flux of 3 × 10−18 erg s−1 cm−2 were
detected. Although the result is impressive in that next to supernovae this discovery is
the most distant measurement of individual stars (100 Mpc), a fair comparison with real
36 Martin M. Roth
Figure 1.45. Multi-Slit-Mask for FOCAS (from Gerhard et al., 2005).
s''
S'
Ω' Ω'' Ω∗
Ω
AS Entrance
A
slit
Figure 1.46. Definition of Etendue of a Spectrograph.
IFUs shows that the uncertain location of point sources over the slit perpendicular to
the slit length obviously introduces large flux uncertainties, making this technique less
competitive compared to, for example, the MUSE instrument.
Nod-shuffle 3D spectroscopy
The technique of ‘nod-shuffle’ spectroscopy is not necessarily restricted to 3D; however, it
is a very powerful option for background-limited observations of extended objects, where
no patch of empty sky can be found within the FoV of the IFU. As described by Cuillandre
et al. (1994), consecutive beam switching between object and sky pointings while shuffling
charge on the CCD detector between an object and a sky integration zone, and final
read-out after the total exposure time has been completed, is capable of eliminating
systematic sky subtraction errors in the sense of chopping technique, which are common
place in many areas of measuring techniques. A comprehensive discussion is presented
by Glazebrook and Bland-Hawthorn (2001), and experience from an implementation in
the PMAS instrument was reported by Roth et al. (2004a).
1.4.8 Figure-of-merit
In order to be in a position to objectively compare competing instrument designs, it is
useful to introduce a quantitative measure of performance. The classical measure to this
end is ‘etendue’, which is defined as:
Etendue = area × solid angle × system throughput (1.4)
Another measure, which takes out the system throughput and thus gauges the perfor-
mance of an IFU irrespective of the attached spectrograph is ‘grasp’:
grasp = area × solid angle (1.5)
Examples for grasp of a number of 3D instruments are presented by Bershady et al.
(2004). Another figure-of-merit, called ‘specific information density’ was proposed by
Allington-Smith (2007).
While these numbers are useful for planning the optimal design of an IFU for a given
science case, observers may also want to consider other parameters, which may be even
more relevant than a single figure-of-merit; for example, spatial resolution, spectral res-
olution, wavelength regime and wavelength coverage.
1.5 Acknowledgments
The Euro3D Research Training Network was funded by the European Commission in
Framework Programme 5 under contract number CT-2002-00305. The initiative of the
IAC, La Laguna, Tenerife, to organize the XVIIth IAC Winter School on the topic of ‘3D
Spectroscopy’ as a task of the Euro3D Work Plan, to provide funding, and to arrange a
marvelous school at the beautiful location of Puerto de la Cruz is gratefully acknowledged.
REFERENCES
Afanasiev, V.L., Vlasiuk, V.V., Dodonov, S.N., Silchenko, O.K. (1990), Preprint No.54,
Special Astrophysical Observatory, Selentchuk
Allington-Smith, J.R. (2007), in Science Perspectives for 3D Spectroscopy (ESO Astrophysics
Symposia), eds. Kissler-Patig, M., Wals, J.R., and Roth, M.M. Springer
Allington-Smith, J., Murray, G., Content, R. ET AL. (2002), PASP, 114, 892
Arribas, S., Mediavilla, E., Rasilla, J.L. (1991), ApJ, 369, 260
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B. ET AL. (1995), A&AS, 136, 189
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B., del Burgo, C., Fuensalida, J.J.
(1998a), A&AS, 136, 189
Arribas, S., Cavaller, L., Garcı́a-Lorenzo, B., Garcı́a-Marı́n, A., Herreros, J.M.,
Mediavilla, E., Pi, M., del Burgo, C., Fuentes, J., Rasilla, J.L., Sosa, N. (1998b),
in Fiber Optics in Astronomy III, eds. S. Arribas, E. Mediavilla, F. Watson, ASPCS, 152,
p. 149
Arribas, S. ET AL. (2005), Bulletin of the American Astronomical Society, 37, 1352
Atherton, P.D., Reay, N.K., Ring, J.R., Hicks, T.R. (1981), Opt En, 20 (6), 806
Atherton, P.D., Taylor, K., Pike, C.D., Harmer, C.F.W., Parker, R.N. (1982), MNRAS,
201, 661
Bacon, R., Emsellem, E., Monnet, G., Nieto, J.L. (1994), A&A, 281, 691
Bacon, R., Adam, G., Baranne, A., Courtès, G., Dubet, D., Dubois, J.P., Emsellem,
E., Ferruit, P., Georgelin, Y., Monnet, G., Pecontal, E., Rousset, A., Say, F.
(1995), A&AS, 113, 347
Bacon, R., Copin, Y., Monnet, G. ET AL. (2001), MNRAS, 326, 23
Bacon, R., Adam, G., Cabrit, S. ET AL. (2002), in Scientific Drivers for ESO Future
VLT/VLTI Instrumentation, eds. J. Bergeron, G. Monnet, Springer, Berlin, p. 108
Bacon, R., Bauer, S.-M., Bower, R. ET AL. (2004), SPIE, 5492, 1145
Balick, B., Preston, H.L., Icke, V. (1987), AJ, 94, 1641
Barden, S., Scott, K. (1986), BAAS, 18, 951
Barden, S.C., Wade, R.A. (1988), ASPCS 3: Fiber Optics in Astronomy, 113
Barwig, H., Schoembs, R., Buckenmayer, C. (1987), A&A, 175, 327
Baudrand, J., Walker, G.A.H. (2001), PASP, 113, 851
Becker, T., Fabrika, S., Roth, M.M. (2004), AN, 325, 155
Bershady, M.A., Andersen, D.R., Harker, J., Ramsey, L.W., Verheijen, M.A.W.
(2004), PASP, 116, 565
38 Martin M. Roth
Bingham, R.G., Gellatly, D.W., Jenkins, C. ET AL. (1994), SPIE, 2198, p. 56
Bland–Hawthorn, J., Taylor, K. (1994), AAO Newsletter, 71, 8
Bohlender (1994), CFHT FTS Manual
Bower, R.G., Morris, S.L., Bacon, R. ET AL. (2004), MNRAS, 351, 63
Buzzoni, B., Delabre, B., Dekker, H. ET AL. (1984), Msngr, 38, 9
Cappellari, M., Copin, Y. (2003), MNRAS, 342, 345
Compte, G., Marcelin, M. (1995), Tridimensional Optical Methods in Astrophysics, N◦ 149.
IAU Colloquium, 1994
Courtès, G. (1960), AnAp, 23, 96
Courtès, G., Georgelin, Y., Bacon, R., Monnet, G., Boulesteix, J. (1988), in Instru-
mentation for Ground-Based Optical Astronomy, Proc. 6. Santa Cruz Summer Workshop,
ed. L.B. Robinson, Springer, New York, p. 266
Cuillandre, J.C., Fort, B., Picat, J.P. ET AL. (1994), A&A, 281, 603
Dean, A. (2002), Thesis, IoA University of Cambridge
de Zeeuw, P.T., Bureau, M., Emsellem, E., Bacon, R., Carollo, C.M., Copin,
Y., Davies, R.L., Kuntschner, H., Miller, B.W., Monnet, G., Peletier, R.F.,
Verolme, E.K. (2002), MNRAS, 329, 513
Eisenhauer, F., Abuter, R., Bickert, K. ET AL. (2003), SPIE, 4841, 1548
Emsellem, E., Bacon, R., Monnet, G., Poulain, P. ET AL. (1996), A&A, 312, 777
Fabrika, S., Sholukhova, O., Becker, T., Afanasiev, V., Roth, M., Sánchez, S.F.
(2005), A&A, 437, 217
Filippenko, A.V. (1982), PASP, 94, 715
Fruchter, A.S., Hook, R.N. (2002), PASP, 114, 144
Gerhard, O., Arnaboldi, M., Freeman, K. ET AL. (2005), ApJ, 621, L93
Glazebrook, K., Bland-Hawthorn, J. (2001), PASP, 113, 197
Hill, J.M. (1988), ASPCS 3: Fiber Optics in Astronomy, 3, 77
Hill, J.M., Angel, J.R.P., Scott, J.S., Lindley, D., Hintzen, P. (1980), ApJ, 242, L69
Hook, R.N., Lucy, L.B. (1993), in Proc. Science with the HST, p. 245
Jacoby, G.H., Kaler, J.B. (1993), ApJ, 417, 209
Jerram, P., Pool, P.J., Bell, R. ET AL. (2001), SPIE, 4306, 178
Kelz, A., Verheijan, M.A.W., Roth, M.M. ET AL. (2006), PASP, 118, 129
Kenworthy, M.A., Parry, I.R., Taylor, K. (2001), PASP, 113, 215
Kissler-Patig, M., Copin, Y., Ferruit, P., Pecontal-Rousset, A., Roth, M.M. (2004),
AN, 325, 159
Larkin, J., Barczys, M., Krabbe, A. ET AL. (2006), New Astronomy Review, 50, 362
Lee, D., Haynes, R., Ren, D., Allington-Smith, J. (2001), PASP, 113, 1406
Le Fevre, O., Saisse, M., Mancini, D. ET AL. (2003), SPIE, 4841, p. 1670
Mediavilla, E, Arribas, S. (1993), Natur, 365, 420
Meisenheimer, K., Hippelein, H. (1992), A&A, 264s, 455
Michlovic, J. (1972), App Opt, 11, 490
Murphy, T.W., Jr., Matthews, K., Soifer, B.T. (1999), PASP, 111, 1176
Ohtani, H., Sasaki, M., Aoki, K., Takano, E., Kiyohara, M. (1994), SPIE, 2198, 229
Parry, I.R., Carrasco, E. (1990), SPIE, 1235, 702
Parry, I.R., Bunker, A., Dean, A. ET AL. (2004), SPIE, 5492, 1135
Perryman, M.A.C., Favata, F., Peacock, A. ET AL. (1999), A&A, 346, L30
Ramsay Howatt, S.K., Todd, S., Leggett, S. ET AL. (2004), SPIE, 5492, p. 1160
Roth, M.M., Becker, T., Schmoll, J. (2000), ASPCS 195: Imaging the Universe in Three
Dimensions, 122
Roth, M.M., Fechner, T., Wolter, D., Kelz, A., Becker, T. (2004a), in Scientific Detec-
tors for Astronomy, The Beginning of a New Era, eds. Amico, P.; Beletic, J.W.; Beletic,
J.E., 371
Roth, M.M., Becker, T., Kelz, A., Schmoll, J. (2004b), ApJ, 603, 531
Roth, M.M., Kelz, A., Fechner, T., Hahn, T., Bauer, S., Becker, T., Böhm, P., Chris-
tensen, L., Dionies, F., Paschke, J., Popow, E., Wolter, D. (2005), PASP, 117, 620
Sánchez, S.F. (2004), AN, 325 (2), 167
Sánchez, S.F., Becker, T., Kelz, A. (2004a), AN, 325 (2), 171
Sánchez, S.F., Garcı́a-Lorenzo, B., Mediavilla, E. ET AL. (2004b), ApJ, 615, 156
Schmoll, J., Roth, M.M., Laux, U. (2003), PASP, 115, 854
Si’lchenko, O., Afanasiev, V. (2000), A&A, 364, 479
Tully, R.B. (1974a), ApJS, 27, 415
Tully, R.B. (1974b), ApJS, 27, 437
Tully, R.B. (1974c), ApJS, 27, 449
Vanderriest, C. (1980), PASP, 92, 858
Vanderriest, C. (1995), in Tridimensional Optical Spectroscopic Methods in Astrophysics, IAU
Coll. 149, ASPCS 71, p. 209
Vanderriest, C., Lemonnier, J.-P. (1988), in Instrumentation for Ground-Based Optical
Astronomy, Proc. 6. Santa Cruz Summer Workshop, ed. L.B. Robinson, Springer, New
York, 304
Walsh, J.R., Roth, M.M. (2002), Msngr, 109, 54
Weitzel, L., Krabbe, A., Kroker, H., Thatte, N., Tacconi-Garman, L.E., Cameron,
M., Genzel, R. (1996), A&AS, 119, 531
Wells, D.C., Greisen, E.W., Harten, R.H. (1981), A&AS, 44, 363
Wilkinson, A. Sharpless, R.M., Fosbury, A.E.M., Wallace, P.T. (1986), MNRAS, 218,
297
Williams, R.E., Blacker, B., Dickinson, M. ET AL. (1996), AJ, 112, 1335
2. Observational procedures
and data reduction
JAMES E.H. TURNER
2.1 Introduction
In this chapter, I give an introduction to observing with integral field units and performing
basic reduction of the resulting data, prior to scientific analysis. After briefly considering
the context of the lectures, I begin by discussing strategies for observing. This is followed
by a short tutorial on sampling theory and its application to integral field unit (IFU) data,
before continuing with an overview of the requirements for each stage of data reduction.
I finish by considering the data reduction process as a whole, along with associated issues
such as error propagation and file formats.
2.2 Background
Techniques for integral field spectroscopy (IFS) have been in development for at least
two decades (Vanderriest, 1980).1 During the 1980s–1990s, numerous prototype IFUs
and even a few public instruments were deployed at observatories and used for scientific
work. Nevertheless, IFS has only become widely available at major telescopes during the
past five years or so, following two centuries of slit spectroscopy. Experience in observing
with IFUs and processing the data is just starting to become commonplace within the
community, but will be spread more widely by the current generation of postdoctoral
and student astronomers.
In terms of data reduction and analysis, IFS poses some non-trivial new requirements.
The most obvious factor is the introduction of 3D datasets to mainstream optical and
near-infrared (NIR) (as opposed to radio) astronomy. Although older scanning methods
such as Fabry–Perot interferometry produce higher-dimensional datasets, these tech-
niques are relatively specialized by comparison. IFS is an extension of traditional spec-
troscopy, covering a range of spatial and spectral sampling regimes, increasingly so as
larger-format units become available. This creates a need for non-specialists to apply
spectroscopic and image processing to 3D datasets. Perhaps a more important issue than
dimensionality, however, is that IFUs scramble spatial information in the process of map-
ping three dimensions onto a 2D detector. To add to this, the spatial pixels of an IFU
rarely form a square grid to begin with. Special software is therefore needed to recover
spatial or 3D images. Depending on the processing strategy, this may affect multiple data
reduction stages. Finally, a third consideration in dealing with IFU data is homogene-
ity, given variations in transmission, spectral resolution or background counts across the
field. This leads to additional flat-fielding requirements, for example, compared with a
simple aperture.
Although instruments with IFUs have become more common in the past few years,
the availability of matching general-purpose software has lagged behind somewhat. A
key task of the Euro3D Research Training Network (2002–05) has been to create data
visualization and analysis tools to help address this issue, as well as a standard file format
for working with reduced IFU data. Other recent progress includes adaptation of image-
viewing programs to work with 3D data. The proliferation of high-level environments such
1
Ignoring the even earlier use of image slicers to avoid slit losses (Bowen, 1938).
40
Observational procedures and data reduction 41
Figure 2.1. Sketch of the relative atmospheric transmission as a function of wavelength in the
optical and near infrared. Some common photometric bandpasses are labelled with the appropri-
ate letters. (Based on a figure from a US Naval publication, reproduced at http://wikipedia.org.)
as Python and interactive data language (IDL), with multi-dimensional array handling
capabilities, also provides some flexibility for manipulating 3D datasets. There is still
progress to be made, however, before the same range of tools exist for working with IFU
data as for slit spectroscopy.
2.3 Observing strategies

2.3.1 IFUs versus other instruments
Current IFUs work at optical wavelengths (∼0.4–1.0 µm, with CCD, charge-coupled
device, detectors) or in the near-infrared (∼1.0–2.5 µm with HgCdTe or ∼1–5 µm with
InSb arrays). Future mid- and far-infrared IFUs are also planned, respectively, for the
James Webb Space Telescope (JWST) (Wright et al., 2004) and Sofia (Raab et al., 2004).
Figure 2.1 is a rough sketch of the atmospheric transmission over the range of wavelengths
covered by existing IFUs, showing the relation to common photometric bandpasses.
The details of how an IFU is used depend on its integration with the host instrument,
as well as on wavelength and the type of IFU. Some integral field spectrographs are
implemented as stand-alone instruments (e.g. Bacon et al., 2001), whereas others are
general-purpose spectrographs with an IFU module that can be inserted in place of an
ordinary slit mask (e.g. Allington-Smith et al., 2007; Figure 2.2 herein). Sometimes an
IFU can be retro-fitted to feed an existing long slit, typically using optical fibres. A hybrid
of multi-object and integral-field spectrograph designs involves individually deploying
multiple small IFU fields within the telescope field; several such instruments have been
proposed and one is already in routine operation (Pasquini et al., 2002). In addition, some
‘single-object’ optical spectrographs include a second IFU field or separate sky fibres for
sampling the background at the same time as a target object (e.g. Allington-Smith et al.,
2002).
Different types of IFUs have a few common characteristics compared with other instru-
ments. In particular, there is a trade-off between the numbers of spatial and spectral
pixels that will fit on the detector (these may also be limited by optical size constraints).
42 James E.H. Turner
Figure 2.2. The image slicing IFU of the GNIRS spectrograph at Gemini South (right),
mounted alongside various slit masks in a motorized sliding mechanism that inserts the appro-
priate mask into the beam.
More often than not, this translates into small field sizes (arcseconds) compared with
typical long slits or imagers (arcminutes or even degrees). Some IFUs, on the other hand,
are designed with coarse spatial pixels and/or short spectra, giving a larger field (∼tens
of arcseconds) and potentially much better sensitivity for low surface brightness objects,
instead of resolving fine spatial detail (e.g. Bershady et al., 2004). Sometimes it is possible
to choose between multiple spatial scales.
A second common characteristic of IFUs is that they introduce non-uniformities over
the field, as mentioned in Section 2.2, especially with optical fibres. Together with a small
field size or coarse sampling this makes dithering and mosaicking important strategies
for observing with IFUs, both to average over variations and to cover a larger area
or improve the sampling (see Section 2.3.2 Dithering and mosaicking, and Section 2.4).
Moreover, the IFU field response has to be ‘flat fielded’, in addition to that of the detector,
unlike the step response of a slit.
Finally, it is worth noting that high-spatial-resolution IFUs are a natural companion for
adaptive optics systems, both scientifically (many spectroscopic projects using adaptive
optics benefit from 2D spatial coverage) and technically (the difficulty of using a slit at
such small spatial tolerances). This is another source of special requirements for observing
and data reduction, albeit outside the scope of the current chapter.
2.3.2 The observing process

Here and in Section 2.5, I shall concentrate on the case of a single target field, observed at
optical or near-infrared wavelengths, since this applies to nearly all current integral field
spectrographs. Many of the points discussed are generic issues in spectroscopy, considered
here in the context of IFS for completeness. Precise details vary significantly between
instruments, with factors such as wavelength, flexure, repeatability and the available
modes or configurations; thus much of the following text is general discussion relating to
each observing or data reduction stage, but Section 2.3.2 (Overview) and Section 2.6 tie
the steps together with examples of specific sequences.
Nod to sk Nod to sk
Ac uire target bserve bserve science bserve bserve science

onto the IF flat and/or arc target xN flat and/or arc target xN
ffsets on source
Figure 2.3. Sketch of a typical observing sequence at optical wavelengths.
bserve bserve
blank sk blank sk
xN xN
Ac uire target bserve science bserve science

bserve flats bserve flats
onto the IF target target
ffsets on source
Ac uire telluric standard bserve standard

bserve flats
star onto the IF at position A
xN 1-4
bserve standard
at position B
Figure 2.4. Sketch of a typical observing sequence at near-infrared wavelengths.
Overview
Figure 2.3 is a sketch of a typical on-sky observing procedure at optical wavelengths. After
centring the target on the IFU, some number of exposures of its spectrum are taken, often
interspersed with flat and/or arc lamp calibration spectra (where flexure is important). If
the target is extended enough to cover the whole field, it may be necessary to observe it
alternately with blank sky; however, this is usually avoidable with separate fibres or lenses
at some distance from the field centre, sampling the background contemporaneously
with the object. Smaller spatial offsets between target exposures are sometimes used for
mosaicking or averaging over flat-field variations.
Figure 2.4 is a sketch of a typical on-sky observing procedure in the near-infrared.
Unlike in the optical case, it is usual to alternate exposures on the science target with
blank sky spectra (perhaps in an ABBA sequence, to reduce offsetting overheads) in order
to sample the object and background with the same optical path and detector pixels.
Given the abundance of telluric emission lines in the near infrared, there is typically less
need to take arc lamp spectra for wavelength calibration, but frequent flat fields may
still be required in the presence of instrumental flexure. Usually a separate standard
star observation is taken – close to the science target in both airmass and time – to
calibrate the telluric absorption features found at longer (red to infrared) wavelengths.
Table 2.1. Typical ‘off-line’ calibration data
Day time or twilight

(with a dark dome) Twilight Night time
Lamp flats Twilight flats Flux standard

Arcs (mainly optical)
Biases (optical) Radial velocity
Darks (mainly IR) or line index standards
Special calibrations
Being compact, the star can normally be offset between two or more positions within the
IFU field for sky subtraction, without spending time off source.
Whatever the observing wavelength, after some number of hours (depending on flexure
or telescope guiding), a re-acquisition step may be necessary in order to maintain good
centring on the target. This is more important if the telescope guider does not compensate
for changing atmospheric dispersion between the observing and guiding wavelengths.
Nevertheless, centring is likely to be less critical than for a typical narrow slit.
More often than not, some additional calibration data are required that need not be
taken on the sky with the science target (Table 2.1). This may include flat-field or arc-
lamp frames if flexure does not make the illumination dependent on time or telescope
pointing. For optical CCDs, bias exposures are needed, to find the zero point of the data.
For infrared arrays or older CCDs, dark exposures may be used to remove counts due
to thermally excited electrons. Twilight sky flats provide the spatially flattest source of
strong illumination if artificial sources are not adequate alone. If absolute flux calibration
is not needed, a spectrophotometric flux standard star can be observed within a reason-
able time of the science data (days or even weeks) to calibrate the sensitivity spectrum of
the telescope+instrument. (In the near infrared, the telluric standard may also serve this
purpose.) Depending on the project, additional stellar observations may be needed as
spectral templates for velocity or line index measurements. Either kind of standards may
be taken in poor observing conditions if absolute fluxes are not required. Finally, special
calibrations are required for some instruments, for example to calibrate the alignment of
the IFU with the detector.
Acquisition
Blind telescope pointing is rarely accurate enough to centre a target perfectly within a
small IFU field, so acquisition often involves iterative centring at the required position –
usually the middle of the IFU – as for a slit. At each step, an exposure is taken, the
target position is measured and the telescope (or in some cases the IFU) is offset slightly
to adjust the pointing.
Measurement of the target position can be done either with a normal imaging camera
that is fixed with respect to the IFU (perhaps a different mode of the same instrument)
or by reconstructing a 2D image of what the IFU itself is looking at (Figure 2.5). In the
former case, the target is offset until it reaches pixel coordinates known to correspond
to the IFU centre, based on prior measurements. This requires that the position of the
IFU on the camera is repeatable, without too much flexure or other sources of significant
error. In the latter case, the best sensitivity is achieved by taking images undispersed
through the IFU (or, if applicable, using the wavelength of an emission line). The 1D slit
or micropupil spots are then rearranged to form a 2D image. If saturation is a concern
Figure 2.5. a) Acquiring a target object into a box marking the IFU position in the GNIRS
direct
imaging field; b) section of an image of the same target, taken undispersed through the IFU;
c) a reconstruction of the IFU field, made by stacking the image slices from the exposure in the
previous panel.
or the spectrograph does not allow taking undispersed images, spectra can be summed
over wavelength before reconstructing a 2D image, or a 3D datacube can be constructed
first and then collapsed in wavelength (if not too slow). Finally, if flexure is of concern
and initial pointing is poorer than a few arcseconds – or the target is faint or difficult
to identify – the two approaches can be combined by using a direct camera image to
place the target within the IFU field first and then fine-tuning the pointing by taking
exposures through the IFU.
When combining exposures from an instrument with large spatial pixels it may help to
have repeat acquisitions (e.g. from different nights) accurately aligned; if spatial informa-
tion is missing due to the coarse pixel size, co-alignment of data during later processing
is likely to be problematic (Section 2.4). Differences in atmospheric dispersion between
exposures, however, tend to limit how well alignment can be maintained over a range of
wavelengths, so some error is often unavoidable. For spatially well-sampled IFU data, on
the other hand, one can correct for both pointing offsets and relative changes in atmo-
spheric dispersion afterwards; thus, in many cases periodic re-acquisition is less important
than for a slit, where miscentring on the aperture is difficult to correct afterwards and
can lead to loss of signal.
Object and sky spectra

Once the target is at the right place in the IFU field, the observer can keep taking
spectra until the required signal-to-noise level is reached (or some practicality interrupts
the observation). As well as integrating on the target itself, one normally has to observe
blank sky as a reference for subtracting out telluric emission lines and other sources of
background counts. Given that IFU dimensions are typically smaller than slit lengths,
the probability of having to nod completely off the target to obtain blank sky is greater.
Figure 2.6. Spectra from separate, interleaved blocks of object and sky fibres in the GMOS
IFU (this detector subsection overlaps 2 of a total of 30 blocks of fibres).
At optical wavelengths, the sky can usually be sampled at the same time as the object,
using one of several methods. If the target is compact enough, one can use blank sky from
the edges of the IFU or nod up and down the field in order to subtract pairs of frames.
With multiple targets one might dither around the field and later co-add exposures
with rejection to remove the objects. Often, however, the instrument is designed with a
separate background field or sky fibres to take care of the problem. These are placed far
enough away from the main science field (e.g. a few arcminutes away) to fall on blank
sky in most cases, or at least to contain negligible signal from the outer regions of an
extended target, compared with the science field. If the target is particularly extended,
however (or maybe the sky field is contaminated by an object), it is necessary to nod
off the science target completely from time to time and expose on blank sky elsewhere.
Figure 2.6 shows an example of separate object and sky spectra from a fibre IFU.
In the near infrared, standard practice is to nod to sky, usually every other exposure, to
allow removal of telluric and thermal emission and detector dark current by subtracting
pairs of raw exposures. Since sky lines are stronger than in the optical and vary on
timescales as short as a few minutes, it is usual to nod to sky more frequently for optimal
subtraction. For compact targets such as standard stars, we can again nod within the
IFU field to subtract pairs of exposures without spending time off source, but more often
than not half of the observing time is spent on blank sky.
At non-thermal wavelengths (∼1 µm) and high enough spectral resolution, it may be
sufficient for some projects to spend all of the observing time on source and later mask
out or interpolate over sky lines, after subtracting separate dark frames instead of sky.
Integration times
IFS exposure times are mainly determined by the same factors as for other spectroscopic
modes. For faint targets, the minimum exposure is, as always, limited by the need to inte-
grate long enough for the background noise to overcome detector read-out noise. With
IFUs it often takes longer to reach the same number of counts per pixel as for a slit.
One reason is that high spatial or spectral resolution IFUs have smaller apertures than a
typical slit, being able to sample finer detail without losing light overall (but, of course,
spreading the light over more apertures ultimately leads to higher read noise). IFUs also
place extra optics in the beam, with a corresponding loss of efficiency (although through-
puts of up to ∼ 90% have been achieved, compared with a slit in the same spectrograph).
A third factor is that IFU optics sometimes introduce extra magnification, which again
means spreading the light over more pixels. On the other hand, IFUs with large spa-
tial elements – designed for high sensitivity rather than fine sampling – can reach the
read-noise limit very much more quickly. For very bright objects, the minimum exposure
time is sometimes set by instrumental factors such as detector read-out speed.
Maximum exposure times are affected by several factors. In the infrared we have to
start a new exposure often enough to allow for variations in sky lines (on multiple charac-
teristic timescales) for accurate subtraction. For bright targets, one can take short expo-
sures of a few seconds to a few minutes, to sample the fast telluric variations. For fainter
targets, longer exposures of 10–30 minutes are used, to average over short-timescale vari-
ations without going so long as to be affected by slower changes. Exposures much longer
than these usually lead to poor sky subtraction, although the stability varies from night
to night. When working between sky lines at higher resolutions, the subtraction accuracy
may be less important. Another common time limitation for bright objects (or perhaps
bright sky lines) is the need to avoid reaching non-linear count levels or saturating the
detector capacity altogether with too many photo-electrons. Finally, there are various
other practical reasons for dividing up observations, such as allowing for changes in flex-
ure between the slit and detector or the telescope image and IFU, the use of repeated
samples to help remove cosmic rays, spatial dithering on the source or simply avoiding
excessive time loss in the event that something goes wrong with an integration.
Given the above considerations, typical spectroscopic exposure lengths for scientific
targets range from a few minutes up to 60 minutes (in the optical) or ∼ 20 minutes (in
the near infrared).
Dithering and mosaicking

There are, broadly, three scenarios for offsetting the position of an IFU with respect to
the target between exposures; I shall refer to them here as dithering, mosaicking and
sub-pixel dithering.2 These are usually carried out by moving the telescope (or tilting its
optics) slightly, but occasionally it is possible to move the IFU itself within the telescope
field. In any of the three scenarios, an accurate way of registering the relative shifts is
needed in order to combine data correctly afterwards. If the system is known to apply
(or report) offsets precisely enough, the nominal shifts can be applied blindly during
data reduction, possibly in addition to wavelength-dependent atmospheric dispersion
corrections. Otherwise one or more image features must be used for reference, to measure
the alignment afterwards.
Dithering involves making small spatial offsets (e.g. 1–2 IFU elements), to average over
instrumental variations. This is useful because IFU optics introduce artificial structure
into the focal plane compared with a clear slit aperture (although even slits can have
small non-uniformities in width). It also helps with detector defects, as for other observ-
ing modes. Whilst flat-fielding corrects throughput differences across the field, dithering
evens out the resulting noise variations and ‘fills in’ any dead elements, especially for
fibre IFUs. It may also smooth over any differences in spectral-line profiles from one IFU
element to another. For short exposures, restricting offsets to an integer number of spatial
pixels sometimes allows multiple spectra to be co-added without interpolation; this only
applies if atmospheric dispersion is similar between the exposures, the IFU grid is regular
and offsetting errors are small, but can help avoid problems due to undersampling (see
Section 2.4).
Mosaicking uses offsets comparable to the size of the IFU field (especially for fields
of only a few arcseconds) to increase spatial coverage. For mosaics of up to ∼ 4 times
the native field of view, the most reliable approach is often to keep the centre of the
target (e.g. galaxy nucleus) within the field of view at every pointing, as a reference for
2
Sub-pixel dithering is referred to as ‘subsampling’ in the original lectures, but that was
perhaps not the best choice of terminology, given prior use of ‘supersampling’ for the same idea.
aligning exposures. If larger mosaics are needed, one either has to rely on the hardware for
accurate movements or use multiple reference peaks with well-known relative positions
to calculate the offsets.
For IFUs with coarse spatial pixels, sub-pixel dithering may help avoid loss of spatial
information and the resulting processing artifacts. This idea is best known in connec-
tion with the Drizzle technique for HST Wide-Field Planetary Camera (WFPC) data
(Fruchter and Hook, 2002). A likely example would be stepping by one half or one
third of a fibre, lenslet or image slice between exposures (much smaller increments do
not necessarily gain extra detail, given their size relative to the smoothing effect of the
larger pixel area, but may improve stability when combining data with offset errors; see
e.g. Lauer, 1999). It is important to know the offsets accurately in order to interleave
exposures correctly. Unlike the Hubble Space Telescope (HST), ground-based observa-
tories suffer from variable seeing and cloud, which are likely to limit how accurately
a higher-resolution image can be reconstructed from multiple poorly sampled datasets.
To the author’s knowledge, sub-pixel dithering has not yet been well tested with IFUs,
although a 3D-capable version of the Drizzle algorithm has recently been written by
National Optical Astronomy Observatory (NOAO) and is being tested for use with the
Near Infrared Integral Field Spectrograph Integral Field Unit (NIFS IFU) on Gemini
North.
At near-infrared wavelengths, dithering and mosaicking could potentially allow spend-
ing a larger fraction of observing time on source than for a single target pointing plus
sky. The most common observing sequence for nodding to sky is probably sky–object–
object–sky, repeated as many times as necessary. This optimizes the signal-to-noise ratio
for faint targets, since there is equal background noise from sky and object exposures.
When offsetting, however, data are combined with relative shifts, so a single sky exposure
can be subtracted from multiple object exposures and the results co-added without
using the same measurement repeatedly at any given position. Thus a sequence such as
sky–object–object (repeated) can be used. The signal-to-noise ratio is the same as when
using equal numbers of object and sky exposures, but only if data do not need binning
or fitting spatially. This does not, of course, remove the need to observe sky frequently
enough to sample temporal variations. In practice, the most conservative schemes (i.e.
sky–object– or sky–object–object–sky– repeated) tend to give the best results.
Flat-fielding
As for other observing modes, flat-field frames are needed as a reference for correcting
instrumental efficiency variations (with x, y or λ). Since an IFU introduces structure in
the focal plane, there are two components to the overall flat-field response: the usual
detector flat, including pixel-to-pixel variations and other multiplicative features, and an
IFU flat, accounting for differences in transmission between the optical fibres, lenslets
or image slices that make up the IFU, or along the length of slices. Variations over
the IFU can be due to several factors, such as fibre stresses (focal ratio degradation),
alignment variations between optical elements, optical bonding, differences in surface
roughness and diffraction losses. The IFU and detector are often flat-field corrected as
a single system, but can be treated separately if necessary due to flexure. In order to
have the correct spatial and spectral characteristics simultaneously, a final flat can be
constructed from more than one source (see below). The spatial structure of the flat cor-
responds largely to the IFU, whilst spectral variations are mainly due to the detector and
spectrograph.
Spectrally, (detector) flats require a dispersed source of illumination with a smooth
continuum: usually a bright calibration lamp. Dispersing the light mimics science data,
a) b)
c) d)
Figure 2.7. Some example IFU flats: a) a calibration lamp spectrum from the GNIRS image
slicer; b) a twilight sky spectrum, showing telluric structure; c) small subsection of a lamp
spectrum from the GMOS fibre IFU, covering two blocks of fibres with a gap between them;
and d) a vertical cross-section of the previous panel, showing the throughput differences and
overlaps between optical fibres.
allowing for any wavelength dependence of the flat-field pattern (e.g. due to quantum
efficiency or fringing). An intrinsically smooth spectrum is needed to distinguish small-
scale flat-field structure from the shape of the spectrum itself, which is typically fitted
and removed during processing. The detector flat may incorporate a fringe pattern due
to internal optical interference, especially for CCDs at longer wavelengths. For the IFU
flat component, calibration lamp data can also be used to measure differences in spectral
response (rather than just overall throughput) between IFU elements, if such differences
happen to be significant. If separate detector and IFU flats are needed, on the other
hand, one could take lamp exposures through a clear long-slit aperture, matching the
spectral setup of the science data without the IFU in the focal plane.
Spatially, a flat illumination source is needed for correcting IFU structure and any
vignetting by the spectrograph. The most uniform reference is the twilight sky (although
the sky has a usable brightness level for an interval of only a few minutes, twice per day;
for high spatial or spectral resolution IFUs, the brightness may even be optimal before
sunset or after sunrise, making for long observing nights). Lighting a blank, neutral
area inside the telescope dome with appropriate lamps is another common technique for
achieving relatively flat illumination. Over the small area of a typical IFU field, however,
even an inbuilt calibration lamp may be sufficiently uniform for the spatial content of
the flat, if designed appropriately. Figure 2.7 shows some examples of different types of
flat-field spectra.
Given that individual IFU spectra are much narrower than for a long-slit, flexure
perpendicular to the dispersion axis is more of an issue; a shift of only a few pixels
could correspond to the width of a fibre/lenslet or a large fraction of an image slice.
It is therefore important that flats are well matched to their corresponding scientific
exposures. Often flexure between the IFU and detector is significant, in which case the
safest strategy is to take frequent flats (e.g. once per hour) between science data, or
at least to take them at the same telescope pointing. As well as matching the detector
illumination accurately, this provides a clear reference for tracing IFU elements in the
science spectra. In the absence of flexure, flats can be taken before or after night-time
observations and applied to data from throughout the night. If some optical element
(such as the tilt of the disperser) moves non repeatably, flats or arcs should be taken
before changing the instrument configuration.
As long as they match the flexure of science data, lamp flats taken through the IFU can
be used to correct both IFU and detector variations together. Where the lamp exposures
are not spatially uniform enough on their own, a separate twilight observation can be
used to apply a flatness correction. For IFU flats that do not match science data, a
detector-only flat should also be obtained, allowing them to be reduced and shifted into
alignment or extracted. Likewise, image slicer twilights may need matching lamp flats to
remove flexure-dependent detector variations that differ from science data.
Wavelength calibration
To determine the wavelength accurately at each detector pixel, a spectrum containing
well-known features at zero redshift is needed, covering the range of the detector. The
features must both be recognizable (by their relative or absolute positions and strengths)
and have accurately catalogued wavelengths. We can then measure their pixel coordinates
and interpolate to find wavelengths at intermediate positions. For IFUs, each separate
1D fibre/lenslet or point in a 2D slitlet spectrum needs calibrating.
Often the best reference source is an arc lamp. Some common examples are argon,
copper-argon, thorium-argon, xenon and krypton lamps, which have many strong lines
with precise laboratory measurements (e.g. Figure 2.8). In the red or near infrared, an
alternative is to use telluric emission lines in the raw science spectra, with the advantage
that flexure is then not an issue. Where this is not an option (such as for high resolutions
or short exposures at thermal wavelengths), well-defined telluric absorption bands can
sometimes be used instead.
Typically, at least one arc exposure is taken in order to derive a detailed wavelength
solution, including non-linear terms. Sometimes this is observed during the day time or
twilight. Where there is flexure, any small zero-point shifts in wavelength can often be
corrected using even a single strong telluric line in the raw science spectra. Sometimes,
however, there are no suitable sky lines at all, especially for high dispersions in the
blue. If wavelength calibration is important in those cases, any flexure must be tracked
using frequent arc spectra between science exposures (ensuring that arcs are taken before
changing the disperser tilt, if it is non-repeatable). Occasionally, multiple arc lamps may
be needed to obtain enough lines over certain short wavebands.
Telluric calibration
At far red to infrared wavelengths, scientific data tend to be affected by telluric absorption
lines. The common photometric bandpasses (e.g. I, z, J, H, K, L, M) are defined in
spectral regions with relatively good atmospheric transmission, but still contain many
minor absorption features due to water vapour and other molecular species. Moreover,
a)
b)
Figure 2.8. a) A subsection of a typical copper-argon arc spectrum, taken with the GMOS
optical fibre IFU; b) an argon arc from the GNIRS near-infrared image slicer.
there is sometimes occasion to observe in between these clean regions, perhaps to measure
a strong emission line that is redshifted from the visible.
For many purposes, telluric absorption left in scientific spectra would be problematic.
Telluric lines can be confused with real stellar features, for example, when using an
automatic algorithm to measure velocities. More importantly, telluric lines can overlap
real spectral features, changing their profiles so that spurious line widths, centres or
strengths would be measured from the raw data.
In order to correct for telluric absorption, a star of known spectral type is observed with
little or no intrinsic absorption at wavelengths of scientific interest; typically, this is an
A-type star (Figure 2.9), which has an intrinsic spectrum that is almost featureless apart
from hydrogen lines. The amount of absorption in telluric bands scales with airmass and
varies in time, particularly with changes in water vapour. Telluric standards are therefore
observed immediately before or after the corresponding science object, with coordinates
chosen to match the airmass, or zenith distance, at which it is observed. Where the range
of airmass variation during a science observation is small (e.g. < 0.3 airmasses), a single
star can be observed, matching the average airmass. For longer observations, the range
of airmass can be bracketed with telluric standard spectra taken before and after.
If the spectral profile of the instrument varies substantially across the IFU field, one
might take telluric spectra with the standard star at multiple positions on the IFU to
attempt to calibrate different regions separately. For image slicers, point-like and diffuse
HD 22789
20,000 21,000 22,000 23,000 24,000

Wavelength (Angstroms)
Figure 2.9. K-band spectrum of an A-type star with telluric absorption lines. With the notable
exception of the Brackett-γ line at 2.17 µm, most of the structure in the spectrum is introduced
by the atmosphere. Normalization is arbitrary.
image features will produce slightly different spectral profiles anyway, having different
shapes across the width of an image slice; this is likely to be a relatively minor source of
error, however, especially after summing over the telluric observation spatially.
Flux calibration
To allow meaningful comparison of fluxes between different wavelengths – or with pre-
vious datasets – a stellar calibration spectrum is normally needed. Often this is just a
relative calibration of the instrumental efficiency variation with wavelength, but where
necessary it may also serve as an absolute throughput or sensitivity measurement for the
instrument+telescope+sky.
Relative flux calibration is the spectral equivalent of flat-fielding on large spatial scales
(whilst flat fielding itself takes care of smaller spectral structure from pixel-to-pixel differ-
ences and fringing). It is needed for taking the ratios of lines at different wavelengths, or
measuring the true continuum slope of a target, for example. Given photometric observ-
ing conditions, flux calibration additionally allows the brightness of a source or spectral
feature to be measured in absolute units (such as W m−2 Å−1 arcsec−2 ). Alternatively,
line strengths can be measured relative to the continuum, as equivalent widths, avoiding
dependence on accurate flux calibration.
To derive the sensitivity spectrum for a set of scientific observations, the reference star
needs to be observed with the same instrument configuration and have a known intrinsic
brightness as a function of wavelength. At visible wavelengths, numerous spectrophoto-
metric standard stars are available, with wavelength-tabulated fluxes in the literature
(e.g. Oke, 1990). In the near infrared, one often just uses a convenient star of well-known
spectral type and with a reliable broad-band magnitude measurement; the intrinsic con-
tinuum can then be modelled with a black-body curve for the appropriate temperature
and magnitude. Figure 2.10 is an example of a near-infrared absolute throughput spec-
trum, derived from a G-type star.
50
40
H-band throughput (%)
30
20
10
0
15,000 15,500 16,000 16,500 17,000 17,500
Wavelength (Angstroms)
Figure 2.10. Absolute throughput spectrum for the NIFS integral-field spectrograph at an H-
band setting; the curve was derived from a G-star observation by comparing the observed counts
with a black-body model of the star’s intrinsic flux.
Figure 2.11. Sketch illustrating the problem of colour-dependent slit losses due to atmospheric
dispersion (differential refraction). Although a stellar image may be well centred on the slit at
the central wavelength (middle circle), the same is not true for the extremal wavelengths unless
the slit is rotated to the parallactic angle, along the direction of atmospheric refraction. This
effect scales with airmass and is strongest at blue wavelengths, in the optical.
Flux standards can be taken one or more times during a given observing run (a short
enough period that the instrument and telescope characteristics will not change sig-
nificantly). They need not be observed at the same time as scientific spectra, unless
absolute spectrophotometry is critical or the same data are to be used for both telluric
and flux correction. Where absolute flux measurements are not required, standards can
be observed through cloud, which to first order is neutral in colour.
Because an IFU normally captures all the light from an incident stellar image, without
slit losses, observing standards is relatively straightforward. Absolute fluxes can be mea-
sured without any special setup and, in principle, a single observation may be used for
both telluric and flux calibration. For long-slit spectroscopy, on the other hand, absolute
flux calibration requires using a special wide slit to capture all the light, resulting in a
lower spectral resolution than for scientific data; since a telluric spectrum must match
the scientific resolution, this has to be taken separately. Moreover, atmospheric disper-
sion causes colour-dependent throughput losses for narrow slits, unless observing at the
parallactic angle (Figure 2.11); thus even relative measurement of a sensitivity spectrum
is easier and more accurate with an IFU.
Detector bias
Raw measurements from a CCD detector include an electronic bias level. Since this
varies slightly over the array, a separate zero point needs to be determined for each
pixel, to allow calculation of the true counts accumulated above the bias. This simply
involves taking a few very short exposures (ideally of zero duration, i.e. immediate read
out), in the dark. The resulting images directly provide the zero-point value, with no
photoelectrons stored, for each pixel. These exposures include read-out noise, of course,
so several are taken and later averaged together, to reduce the noise to a level that will
not degrade scientific data when subtracted. A typical bias frame looks flat, with random
variations of just a few counts superimposed; depending on how well the electronics are
isolated from electrical noise sources, the image may also include some faint structure.
The bias level produced by detector electronics can drift with time, but is usually
stable enough for reference exposures to be taken only occasionally, during the daytime.
Sometimes the detector controller also creates an overscan region for each exposure,
by continuing to read out the detector after shuffling and retrieving all the accumulated
charge. This adds extra blank image columns, for monitoring variations in the overall bias
and read-out noise. If necessary, the overscan can be used to make an overall zero-point
correction, on top of the pixel-to-pixel differences from bias exposures.
Infrared arrays are normally read out by measuring the difference in counts between
the start and end of each exposure. This is possible because each pixel can be read out
rapidly without affecting the stored charge, whereas CCDs are read out destructively, by
shuffling the charge from one column to the next and finally off the detector. Thus, the
bias level is removed automatically for infrared observations and does not need measuring
separately.
Dark current
Over the duration of a long enough exposure, detector pixels accumulate some electrons
due to thermal excitation and array defects, as well as from incident photons. Although
detectors are cooled with nitrogen or helium to minimize this thermal current, cooling
is limited by the need to avoid poor quantum efficiency at very low temperatures. For
modern CCDs, the effect may nevertheless be negligibly low (e.g. 1 e− hour−1 pixel−1 ).
For infrared arrays, on the other hand, dark current is more important and tends to affect
some pixels much more than others. This creates variations in the raw data that obscure
real features (Figure 2.12). A reference is therefore needed for distinguishing dark counts
from those due to photons from a target object.
For any given exposure time that is used in observing, the dark current component
of images can be estimated by exposing the detector in complete darkness for the same
length of time. Multiple exposures are usually taken, as for bias frames, to account for
read noise and to allow removal of cosmic rays. Since, however, the exposures may be
much longer than biases (and may therefore be best taken at night or in twilight if light
leaks are an issue), it is not always practical to observe more than a few frames.
For scientific observations with separate, equal sky exposures, the same dark current
is present in both object and sky data. Thus in the infrared, subtracting object and sky
pairs removes dark current automatically from those spectra, without the need for special
calibrations. This usually does a good job of cleaning up most hot pixels, leaving just a
statistical increase in noise; any pixels that remain affected after subtraction are masked
out during reduction. Separate dark calibrations are required, however, in cases where
background subtraction is attempted without nodding to sky, or where sky frames will
be scaled to match on-target data more accurately.
Figure 2.12. Section of a raw scientific exposure from an HgCdTe array. Most of the numerous
bright specks superimposed on the target and sky spectra are ‘hot’ pixels, with high dark
current.
For infrared calibration data, matching darks are usually required unless the exposures
are short enough not to have significant dark counts. For flat fields of more than a
few seconds, it is common to take pseudo-darks in the same configuration as the flats
themselves, but with the lamp turned off. For arc spectra, darks are needed if uncorrected
hot pixels could otherwise be confused with emission lines in a 1D section.
In summary, darks are often needed at infrared wavelengths for calibrations, but are
optional (and time consuming) for scientific data, depending on the reduction method.
With modern CCDs, darks are often not needed at all (but since separate sky frames are
usually not available as a reference, they would be more critical if required).
Summary of observing strategies

On the whole, observing with a single IFU is not radically different from standard slit
spectroscopy. As usual, different techniques apply at optical and near-infrared wave-
lengths. A few issues stand out, however, compared with other spectroscopic modes.
Since IFUs perform imaging as well as spectroscopy, their use reflects both aspects.
Target acquisition allows 2D image reconstruction through the IFU, for example. Spatial
dithering and mosaicking are also important, involving offsets of varying sizes in both
dimensions. Some other practical differences relate to flat-fielding requirements, sky sub-
traction and data inspection. The IFU itself has to be flat fielded, as well as the detector,
carefully accounting for any flexure. At optical wavelengths, sky subtraction may involve
separate sky fibres or a background field, instead of interpolating between the ends of
a slit. In the near infrared, one typically has to nod off to sky instead of up and down a
slit. When viewing data, one has to learn how to interpret scrambled images and/or use
special image or datacube reconstruction software. Finally, sensitivity is often an issue
for high-resolution IFUs, even on 8 m telescopes (although large-pixel designs can do well
compared with a slit).
2.4 Sampling images
Sampling theory helps us understand the effects of dividing an image or waveform into dis-
crete points, such as detector pixels. This is fundamental in digital imaging, but appears
rarely to be taught or discussed outside the engineering and information-theory com-
munities. In astronomy, one sometimes encounters the assertion that ‘Nyquist sampling’
requires two detector pixels per full-width-at-half-maximum (FWHM) of a point source.
Here I show where this criterion comes from and consider how the results of sampling
theory relate to IFU data reduction.
Sampling theory is particularly relevant to IFS data because of the way in which
images are dissected both by the IFU and the detector. This not only introduces an extra
sampling stage, in effect, but often results in a spatial grid which is non-rectangular,
or even completely irregular. Thus resampling in up to three dimensions is generally
necessary if we want to get data onto a cubic or square grid that is easy to work with.
This is complicated by the fact that IFUs sometimes undersample telescope images with
large spatial pixels, for sensitivity reasons or to maximize the field of view. It is important
to understand the consequences of this last case in order to avoid introducing artifacts
during processing.
2.4.1 The Sampling Theorem

The Sampling Theorem deals specifically with sampling on a regular grid. It is often
named after one or more of the following researchers: Kotel’nikov (1933) of the Rus-
sian Academy of Sciences; Nyquist (1928) and Shannon (1949) of Bell Labs, USA; and
Whittaker (1915) of Edinburgh University, UK. Various naming combinations are used
in the literature, such as ‘Shannon’s Theorem’ or the ‘Shannon–Whittaker–Kotel’nikov
Theorem’. Roughly speaking, the idea is that an analogue signal containing a finite
amount of information can be recovered completely from a large enough set of discrete
samples.
For simplicity, consider a one-dimensional function, f (x). Sampling at a particular
point, x′ , can be represented using a Dirac delta (impulse) function:
! ∞
f (x′ ) = δ(x − x′ ) f (x) dx. (2.1)
−∞
In order to have this property, δ(x) can be any function of unit integral that is ‘infinitely
narrow’, i.e. zero everywhere except at x = 0. Multiplying a function by delta therefore
produces an impulse whose area is normalized to the value of the function at the relevant
point (in this case x′ ):
f (x) δ(x − x′ ) = f (x′ ) δ(x − x′ ). (2.2)
Graphically, we represent this delta function as an arrow whose height, f (x ), indicates
′
its normalization (Figure 2.13).

A comb (or shah) function is an infinite series of equally spaced delta functions:
"x# $ ∞
III = |a| δ(x − an), (2.3)
a n = −∞
(where the multiplicative |a| just preserves the unit area of δ when scaling the coordinate
argument by the same factor). Thus, sampling f (x) at regular intervals of a is equivalent
to multiplication with a comb function:
"x# ∞
$
III f (x) = |a| f (an) δ(x − an). (2.4)
a n = −∞
Figure 2.13. Pictorial representation of sampling with δ(x − x′ ).
f(x)
III (x/a)
III(x/a) f(x)
x
a
Figure 2.14. Sampling a continuous function, f (x), at regular intervals of a, by multiplication
with a comb function.
Clearly, III f is a series of impulses, normalized to the sampled values: it contains the
same information from f (x) as would point-like measurements (Figure 2.14).
The comb function is a useful tool for representing sampling because its Fourier trans-
form is also a comb. Working with frequencies in reciprocal units of x (rather than 2π/x)
gives:
"x# F
III ↽ a III(au).
⇀ (2.5)
a
According to the Convolution Theorem, multiplication of functions in the spatial
domain is equivalent to convolution of their Fourier transforms (and vice versa). Thus
the representation of sampled data in frequency space is:
" "x# #
F III f (x) = a III(au) ∗ F (u). (2.6)
a
This is the frequency spectrum of the original, continuous f (x), replicated infinitely at
intervals of 1/a (Figure 2.15); convolving F (u) with each δ(u − u′ ) simply shifts the origin
to u′ .
A function is said to be ‘band limited’ if its Fourier transform is zero above some cut-
off frequency, uc . In Figure 2.15 and Equation 2.6, the replicas of F (u) do not overlap if
f (x) contains a maximum frequency component of uc ≤ 1/2a; thus we can separate out
F (u) from the transform of the sampling pattern and recover the original f (x) without
any loss of information. If the replicas do overlap, on the other hand, the function cannot
be determined uniquely.
III (au) * F(u)

F(u-1/a)
u
1 uc
a
Figure 2.15. The Fourier spectrum of a uniformly sampled function, III(x/a)f (x), is the Fourier
spectrum of f (x), repeated at regular intervals of 1/a.
III(au) * F(u)
u
Figure 2.16. Given the Fourier transform, III(au) ∗ F (u), of a function sampled above the
Nyquist rate, the transform, F (u), of the original, continuous function can be recovered by
multiplication with a box function, Π(au).
The Sampling Theorem can therefore be stated as follows: ‘A band-limited function

can be recovered completely from a set of regular samples, as long as the sampling rate is
at least twice the highest component frequency.’ This is clearly also equivalent to saying
that the separation between samples must be no greater than half the shortest component
wavelength.
For a given sampling interval, a, the maximum distinguishable frequency of 1/2a is
known as the Nyquist frequency. Likewise, the critical sampling rate of 2uc for a fixed
bandwidth of f (x) is often called the Nyquist rate. Sometimes these terms are inter-
changed, however, or used to refer to uc .
To recover f (x), we can isolate F (u) by multiplication with a box function, Π(au), that
has unit value between ±1/2a and is zero elsewhere (Figure 2.16).3 This has the Fourier
transform sinc(x/a)/a, where sinc(x) = sin(πx)/πx is an infinite function with zeros at
integer x (Figure 2.17). Thus, by the Convolution Theorem, f (x) is reconstructed in the
spatial domain by a summation of regularly spaced sinc functions (Figure 2.18):
" "x# # sin(πx/a)
f (x) = III f (x) ∗ (2.7)
a πx
$ ∞
sin( πa (x − an))
= |a| f (an) . (2.8)
n = −∞
π(x − an)
Having reconstructed the original analogue signal, we can then resample it arbitrarily
onto a convenient grid (such as a regular datacube, for IFU data).
3
The box cut-off can be anywhere between uc and 1/2a (if different), but the latter is
convenient because it is known and fixed for a particular sampling.
Figure 2.17. The sinc function, sin(πx)/πx, which peaks at unit value at x = 0 and oscillates
towards small values at infinity, with zeros at integer x.
Figure 2.18. Reconstruction of a continuous function from regularly spaced samples, by convo-
lution with a sinc function. Each grey curve is a sinc function centred at the appropriate sample
point and normalized to the corresponding sample value (cf. Figure 2.14). The thicker black
curve shows the overall sum as a function of x.
III (au) * F(u)
u
1 uc
a
Figure 2.19. When f (x) is undersampled, such that it contains frequency components > 1/2a
for a sampling interval of a, replicas of F (u) overlap in the Fourier spectrum of the sampled
data.
Equation 2.8 specifies f (x) at all intermediate points between the samples. This is
possible because the number of frequency components was finite to begin with, so f (x)
can be represented by a corresponding number of basis functions. The basis could instead
be a Fourier series, but finding the coefficients would then be more complex than above:
sinc functions are a natural basis, because the required coefficients are simply the sample
values.
2.4.2 Undersampling
For a given sample separation, a, if f (x) contains frequencies greater than the Nyquist
frequency, 1/2a, it is said to be undersampled. In the Fourier spectrum of the sampled sig-
nal, F (u) then overlaps its neighbouring replicas (Figure 2.19); the spectrum is effectively
fn(x)
x
a
Figure 2.20. An undersampled frequency component and its lower frequency alias have the
same values at regular sampling intervals, making them indistinguishable. The Nyquist, or crit-
ical sampling, criterion requires the sampling rate to be high enough with respect to the signal
bandwidth that such ambiguities cannot occur.
wrapped around or reflected at 1/2a, such that power at frequencies above this limit is
aliased to lower frequencies. The aliasing can be thought of as beating between high-
frequency components of the signal and the sampling period itself. On a regular grid,
high- and low-frequency pairs can appear identical at every sample point, even though
their values in between are quite different (Figure 2.20). The only way to distinguish
between these aliases is to disallow one of them (the higher frequency in our case); hence
the requirement for band limitation.
Where undersampling occurs, the transfer of power from higher to lower frequen-
cies creates artifacts in the reconstruction of f (x) from samples. Any super-Nyquist
frequencies are not only unmeasurable, but contaminate otherwise good data. We can
end up with less useful information than if f (x) had been smoothed to a lower resolu-
tion/bandwidth before sampling.
With irregular sampling patterns, the similarity between alias frequencies is broken – as
can be seen from Figure 2.20 – so there is no direct transfer of power from one frequency
to another. Nevertheless, information is still lost if the sampling rate is too low, which
unavoidably leads to reconstruction error. Moreover, if the number of available samples is
too small, the problem of finding a set of basis coefficients may become ill conditioned. In
general, the theory of reconstruction from irregular samples is considerably more involved
than that discussed here (see references in Section 2.4.4, Irregular sampling).
Further information on regular sampling, including proofs of the theorems relied on
here, can be found in Bracewell (2000, or earlier editions).
2.4.3 Sampling images

The Sampling Theorem deals with infinite and conveniently band-limited signals, so the
next step is to relate this to real-world images.
As will be familiar to the reader, when a plane wave passes through an aperture, it is
diffracted, so that the light ray leaving the aperture is spread out in angle (Figure 2.21).
The image produced on a screen or detector by the resulting optical interference pattern
depends on distance from the aperture. In the far field, where the aperture is effectively
a pinhole, the amplitude of the diffraction pattern as a function of angle is given by the
Fourier transform of the function describing the aperture transmission (i.e. shape). What
we actually measure is intensity (I ∝ A2 ), so in practice we are interested in the square
of the Fourier transform. For a rectangular aperture (described by a 2D version of the
box function discussed in Section 2.4.1), this is a product of sinc functions. For a circular
telescope mirror, it is the closely related Airy pattern, shown in Figure 2.22:
% &2
J1 (πDθ/λ)
I(θ) ∝ , (2.9)
πDθ/λ
Figure 2.21. Diffraction of a plane wavefront by a finite aperture. The resulting interference
produces a far-field intensity distribution similar in cross-sectional form to the sketch on the
right-hand side (where intensity increases to the right).
( 2 Jπr(πr))
2
1
1.0
0.5
r
1.22
1.02
Figure 2.22. Cross-section of the Airy intensity pattern produced by diffraction of a plane wave
by a circular aperture.
where D is the mirror diameter, λ is the wavelength of light, θ is the angular displacement
in radians from the central direction of propagation and J1 is a first-order Bessel function
of the first kind. A derivation can be found in Goodman (1968) or Hecht (1987).
The finite image of a point source produced by a given optical system is known as its
point-spread function (PSF). For a diffraction-limited telescope, the PSF is therefore the
Airy pattern described above (ignoring a smaller correction for the central obscuration).
For a seeing-limited telescope, the PSF is a superposition of atmospheric speckle images
that tends towards a Gaussian-like form over long enough (∼ 60 s) exposures.4 Since the
PSF describes the distribution of light from any point-like element, the 2D image formed
by a telescope is equivalent to an infinitely detailed projection of the sky, convolved with
the PSF.
The Fourier power spectrum of the PSF is known as the modulation transfer func-
tion (MTF). This describes the sensitivity of the system to different spatial frequencies,
4
In practice, this form is often better described by a Moffat profile (Moffat, 1969), which is
centrally similar to a Gaussian curve, but allows for broader wings.
a) b)
Figure 2.23. The MTF of a telescope is described by the self-convolution of the transmission
over its aperture (so there is always some high-spatial-frequency cut-off).
convolution with the PSF being equivalent to multiplication by the MTF in the fre-
quency domain. Since a telescope’s diffraction PSF is the squared Fourier transform
of its aperture transmission, the Convolution Theorem also tells us that the MTF is
given by the self-convolution of the aperture transmission (Figure 2.23). It follows that,
for a finite aperture, telescope images must be band-limited, with some high-frequency
cut-off.
For a circular mirror of diameter D, the cut-off frequency turns out to be D/λ cycles
per radian (see Goodman, 1968). To satisfy the Sampling Theorem, the detector pixel
spacing must therefore be less than λ/2D radians. The FWHM of the Airy pattern is
1.02 λ/D radians, so in terms of the PSF the critical pixel spacing is 0.49 × FWHM.
Thus we confirm the two-pixel (really 2.04 pixel) sampling criterion. Strictly speaking,
however, this applies only to diffraction-limited imaging.
For seeing-limited images, we can approximate at least the bright core of the PSF by a
Gaussian curve with a FWHM typically an order of magnitude greater than the diffrac-
tion limit. The Fourier transform of a Gaussian curve is also Gaussian, so is infinitely
extended. Thus seeing-limited images do not have a strict frequency cut-off, other than
that imposed at much smaller spatial scales by the diffraction limit. Theoretically, we
would therefore need a FWHM of something like 20 times the pixel size to reach the
Nyquist rate. Fortunately, in practice, we can approximate critical sampling with a less
stringent criterion.
The Gaussian distribution and its Fourier transform can be written
2 1 −π (u /a) 2
F
e−π (ax) ⇀
↽ e . (2.10)
a
√ '
For unit a, these have a standard deviation of 1/ 2π and a FWHM of 2 (loge 2)/π ≈
0.939. Although both curves are infinitely extended, their values die away quickly after
a few standard deviations. If a Gaussian PSF is imaged with a FWHM of n pixels, then,
rewriting its transform in terms of the FWHM using na = 0.939, we have
2
u2
F (u) = 1.06ne−3.56n , (2.11)
with u in cycles per pixel and a standard deviation of σ = 0.375/n. For a FWHM of
n = 2 pixels, the Nyquist frequency of u = 0.5 cycles/pixel therefore corresponds to 2.7σ,
where the Gaussian MTF is at only 2.8% of its peak value. Super-Nyquist frequencies are
therefore heavily suppressed in amplitude, to < 3%. The total integrated power that is
aliased back into the main spectrum is only 0.7%. Thus, even for seeing-limited images,
the ‘2 pixel per FWHM’ criterion is good enough in most cases. A pixel density slightly
greater than this will reduce aliasing to very low levels.
2.4.4 Practical issues
Finite detector
The above discussion outlines how the Sampling Theorem relates to imaging, but several
idealizations still need justification. Perhaps most obvious is the fact that sampling is
done with ∼ 103 detector pixels or ∼ 101 spatial IFU pixels per dimension, rather than
an infinite comb function.
A finite detector image is equivalent to a comb function multiplied by the telescope
image (as previously) and truncated with a box function of width na, where n is the num-
ber of pixels and a their separation. The Fourier transform is therefore an infinite comb
convolved with F (u) (compare Figure 2.15), convolved again with a narrow sinc function
whose first zeros are at ±1/(na). This zero location corresponds to twice the Nyquist
frequency divided by n (i.e. two pixels of the discrete Fourier transform). The extra
convolution reflects the fact that a finite band-limited image is described by a finite
number of frequency components; otherwise, the transform of the sampled data would
be infinitely detailed. Thus, to first order the finite detector size makes no difference; the
transform of sampled data is still, approximately, F (u) repeated at intervals of 1/a. Con-
volving F (u) with something much narrower than itself does not extend the frequency
band appreciably, so the same sampling criterion applies.
To second order, things are a bit more complicated. The sinc function is infinite and
slowly decaying at lower power levels, so the transform of the sampled data actually has
no hard cut-off frequency. In fact, a finite function, such as a sampled image, can never
be strictly band-limited. This frequency extension is manifested as reconstruction error
towards the edges of the data, similar to the oscillation seen at the ends of a Fourier
series when approximating a sharp edge. The behavior can be improved somewhat by
suitable filtering.
Finite pixel size

Whereas delta functions are infinitely narrow, pixels are as wide as their separation. At
each sample point, the telescope image is averaged over the area of a pixel, which to first
order can be a 2D box function, a hexagonal lenslet or a circular fibre. The measured
sample value is therefore the convolution of the telescope image at the relevant point
with the pixel shape. This just contributes an extra term to the overall measured PSF,
along with the natural seeing, telescope diffraction PSF and instrumental aberrations.
Since the effect is to convolve with the pixel shape prior to sampling, the idea of point-
like samples remains intact; the comb representation and pixel size are mathematically
separable.
Noise
Our discussion of the Sampling Theorem assumed noiseless data, so here we must consider
the effect of random errors. Poisson noise is incoherent and spread over all frequencies,
rather than contributing power at any particular frequency. Noise aliasing is not a con-
cern, because it would just fold noise back into the main spectrum, affecting the statistical
realization of the errors without systematic effect. The question is therefore how the noise
level propagates through the resampling process. When interpolating images, the value
at a particular x is a weighted sum of sinc functions, one per sample (see Equation 2.8
and Figure 2.18). Since noise terms add in quadrature, each of the original sample values
contributes to the total variance at x according to the square of its weight. Since the
values of sinc2 (x) at unit intervals add up to one (Bracewell, 2000), the noise level in
the interpolated image is the same as the original. Noise therefore gets carried through
a) b)
x u
−a a 1 uc
2 2 a
Figure 2.24. a) Nearest-neighbour interpolation uses the closest sample value to each x and is
therefore equivalent to convolving a series of point samples with Π(x/a). b) The corresponding
envelope in the spectral domain is ∝ sinc(au), whose zero values coincide with the sampling
replicas of the original image spectrum, F (u).
the system without systematic effects, to first order. In practice, of course, the noise
level will vary over the image and the sinc function may allow bright peaks to dominate
further out.
Practical interpolants
Although a sinc function is the ‘ideal interpolant’, it is infinite and strongly oscillating, so
is not used in practice without modification. Discontinuities such as edges and bad pixels
can cause severe diffraction-like ‘ringing’ effects (Gibbs’ Phenomenon). Tapering the sinc
function with an envelope such as a Gaussian or cosine multiplier provides a more robust
and computationally feasible solution. This type of interpolant is good for well-sampled
data, albeit relatively slow to compute unless look-up tables are used. It still tends to be
more sensitive to spikes and discontinuities than other interpolants, but the envelope can
be tuned to make a trade-off between suppression of ringing and avoiding smoothing. An
equivalent technique is to take the fast Fourier transform (FFT) of the sampled data,
apply a bandpass filter in the Fourier domain, shift in phase and convert back again.
Nearest-neighbour (zero-degree spline) interpolation is the simplest and fastest to com-
pute method. This is equivalent to convolving the sampled data with a box function,
Π(x/a), instead of sinc(x/a). Since the Fourier transform is a sinc function, it does not
actually cut off F (u) at the Nyquist frequency, although it attenuates higher frequen-
cies (Figure 2.24). The first zeros are at ±1/a, or twice the Nyquist frequency. Beyond
that, sinc(au) goes negative and therefore reverses the phase of some frequencies. The
frequencies passed above 1/2a correspond to the ‘blockiness’ of the nearest-neighbour
interpolant, which can be considered a reconstruction error.
In terms of reproducing the analogue telescope image, the nearest-neighbour algorithm
is about the least accurate method. Nevertheless, the values are guaranteed to be rea-
sonable, since they are the same as the original samples: blockiness is an artifact, but a
well-behaved and understood one. It should also be noted that this interpolant does not
smooth the data, as each value comes from a single sample. Resampling with the same
interval at new points effectively just introduces a small coordinate shift. Since both
the values and gradient are discontinuous, the reconstruction cannot be used directly for
fitting smooth contours or velocity fields (although it might be possible to resample first
at the same rate and then use a smooth interpolant, if the sampling grid is regular).
Linear (first-degree spline) interpolation is equivalent to convolving the sampled data
with a triangle, or tent, function, Λ(x/a) (Figure 2.25). This is the self-convolution of the
box function, Π(x/a), and has width 2a. Its Fourier transform is therefore ∝ sinc2 (au),
which dies away much quicker than sinc(au) in the wings and is positive everywhere,
but still does not reach zero until ±1/a. The frequencies passed above the Nyquist limit
x
−a a
Figure 2.25. Linear interpolation is equivalent to convolving a series of point samples with
Λ(x/a). In this example, the dark line shows the summation of three individual triangle func-
tions, one per sample.
a) b) c)
Figure 2.26. Resampling a linear interpolant at the original rate is equivalent to sampling in
the first place with pixels of a size and sensitivity distribution that depend on the offset. In
each of the three cases above, the height at each point represents the relative contribution to
the interpolated value at the position of the arrow.
correspond to the ‘jaggedness’ or discontinuous gradient of the reconstruction. The inter-

polated values themselves are continuous, however, and follow the original curve more
closely than in the nearest-neighbour case. Linear interpolation is still relatively inaccu-
rate compared with higher-order interpolants, but once again the result is robust and
guaranteed to be ‘reasonable’, since the interpolation is bounded by the original sample
values.
It is instructive to consider what happens when resampling a continuous linear
interpolant (which is what people often think of as ‘interpolating’). Compared with
nearest-neighbour, convolving the data with an additional Π(x/a) term (or multiply-
ing the transform by an extra sinc(au)) suppresses high frequencies more. Depending
where we resample with respect to the original points, the Fourier components effectively
combine with different phases, producing different degrees of smoothing. Resampling at
the original points clearly just reproduces the original values, with no smoothing at all.
Resampling at the mid-points, on the other hand, is the same as averaging pairs of sam-
ples and gives maximal smoothing; this is exactly equivalent to sampling with pixels
that are twice as large to begin with5 , so the values are guaranteed to be sensible even
when undersampled. Finally, resampling at some arbitrary position is equivalent to initial
sampling with pixels that have a spatial step in sensitivity from one half to the other
(Figure 2.26).
Cubic (third-degree) splines are a special case of piece-wise cubic polynomials, where
the coefficients of each piece are constrained to give continuous first and second deriva-
tives at the knots, hence a smooth curve. Following on from the nearest-neighbour and
linear cases, the cubic b-spline kernel is Π(x/a) convolved with itself three times. Its
5
Since the sampling rate is unchanged, the larger pixels would have to overlap, as for
multiple dither positions.
Fourier transform is therefore ∝ sinc4 (au), which quickly drops to ∼ 0 beyond the central
peak, but also suppresses frequencies below the Nyquist limit substantially. In fact, the
Central Limit Theorem (Bracewell, 2000) asserts that the result of repeated convolution
tends towards a Gaussian curve, so higher-order b-splines and their transforms are all
Gaussian-like. Convolving sampled data with a cubic b-spline gives a well-behaved result,
which is more accurate than nearest-neighbour or linear interpolants but which causes
strong smoothing of the data. It is also not strictly an interpolation because it does not
pass through the original samples.
Better results can be achieved using a cubic spline curve that does interpolate the data.
Instead of simply convolving samples with the cubic b-spline kernel, a set of coefficients
can be derived such that the interpolant behaves more like the ideal sinc reconstructor
(e.g. Lehmann et al., 1999). This is equivalent to convolution with a kernel that is similar
to a sinc function, with both positive and negative values but less ‘rippling’ beyond the
central peak. Its Fourier transform is box-like, but with more of a tapered cut-off. In
practice, cubic spline interpolation provides a good approximation to sinc reconstruction,
but with better cut-off characteristics and a reasonable number of values to compute at
each point. It is therefore the best compromise for many applications involving reasonably
well-sampled data.
Polynomial interpolants can also be constructed that directly approximate the form of
a sinc function, for use as convolution kernels. These can again be simple and relatively
compact, producing sinc-like results with less computation and less oscillation than a
summation of sinc functions. Some other possible basis functions include wavelets, Fourier
series, Taylor series and Gaussian summations.
Irregular sampling
It is possible to formulate more general theories for irregular sampling, but the relatively
advanced mathematics is beyond the scope of this lecture (for an overview, see Marvasti,
2001). In principle, a band-limited signal can be specified by as few samples as it contains
unknown frequency coefficients – at least the average interval between samples must
meet the regular Nyquist criterion (e.g. Beutler, 1966). However, the vectors describing
sample weights in the space of the basis functions are no longer orthogonal for non-
uniform spacing, so reconstructing the signal can become an ill-conditioned problem.
For guaranteed stability, some methods require that the maximum separation between
samples is no greater than the Nyquist interval.
Reconstruction from irregular samples can be approached in more than one way. For
specific sampling patterns, an analytical solution can be derived; a simple case is that
of interlaced grids, dealt with by Bracewell (2000). Alternatively, one can approximate
the signal by fitting regular basis functions, perhaps iteratively (e.g. Strohmer, 1993).
Either way, any clustering of samples can lead to strong amplification of noise in the
reconstruction, especially when working close to the Nyquist limit (e.g. Lauer, 1999).
Irregular sampling is most likely to be of concern where the spacing varies between
adjacent elements at the IFU input, particularly with optical fibres (assuming the sam-
ple density is high enough in the first place to allow interpolation). It may also occur
indirectly in the rare case where fibre profiles are undersampled at the detector (Turner,
2001). The spatial pixels of an image slicer form a grid that appears irregular in two or
more dimensions, but which can be separated into a series of regular grids in 1D. For
datasets that are only slightly non-uniform on large scales due to geometric distortions, it
may be acceptable just to treat the sampling as regular, as long as the Nyquist criterion
is still met, and then resample the interpolant at variable intervals corresponding to a
regular grid in the appropriate system.
2.4.5 Summary and application to astronomical data
For well-sampled data, it is possible to resample images with very little degradation, as
if we had sampled with a convenient regular grid to begin with. A diffraction-limited
telescope image can be recovered in its original analogue form as long as we sam-
ple with ≥ 2 pixels across the width of the Airy PSF. For seeing-limited images, at
least 2 pixels/FWHM are needed to reduce aliasing to a level comparable with other
errors, such as flat-fielding systematics or noise. For accurate reconstruction with mini-
mal smoothing, a tapered sinc function or smooth piecewise interpolant such as a cubic
spline is likely to be the best option.
Close to the Nyquist limit or in the presence of image defects, it is best to avoid convo-
lution kernels with too sharp a cut-off in the frequency domain, especially truncated sinc
functions; these accentuate the ringing close to sharp features that occurs when idealiza-
tions such as diffraction-limited resolution are not met. Smooth spline-like interpolants
with compact kernels are better behaved, but still prone to some oscillation. Convolution
kernels that are positive everywhere are insensitive to such issues, but can cause strong
smoothing, especially near the Nyquist limit.
For undersampled data, one only knows the image intensity at the sample points,
with insufficient information to say what happens in between. The ideal strategy is
therefore to avoid interpolating altogether. This is not always practical, however, when
data are on a non-rectangular grid with distortions and perhaps mutual offsets. It may
be necessary to interpolate in order to combine datasets or work with a square grid
(otherwise analysis would require special end-to-end treatment, incorporating basic pro-
cessing). Unfortunately, sinc-like interpolation of undersampled data would introduce
serious aliasing artifacts, whilst smooth, positive convolution kernels could cause exces-
sive loss of detail. The best compromise may therefore be to use simple nearest-neighbour
or linear methods, or the related Drizzle algorithm (Fruchter and Hook, 2002). These are
relatively inaccurate in terms of reconstructing a smooth image, but ensure that values
remain meaningful. They can, however, introduce their own peculiarities in the presence
of image distortions, such as skipping nearest neighbours or variable smoothing.
Whilst the above discussion considers sampling telescope images, the same principles
apply to sampling a dispersed spectrum. The spectral PSF of a slit spectrograph or image
slicer is a rectangular slit (for extended sources) or a narrow section of the telescope PSF,
truncated by the slit (for stars). Although the edges are smoothed slightly by the finite
resolution of the optics, these spectral profiles pass high frequencies more strongly than
a Gaussian or Airy profile of similar width, so 2 pixels/FWHM does not work quite
as well. One may therefore need to be careful to use an interpolant that damps high
frequencies without a sudden cut-off. This is less of an issue for spectra with intrinsically
broad features, where the sampling may be good even though the instrumental PSF for
narrow lines would be undersampled. Fibre or lenslet spectra, on the other hand, often
produce profiles closer to a Gaussian curve, so can be treated more like seeing-limited
images.
2.5 Overview of data reduction issues

The aims of basic IFU data reduction are fairly straightforward. Ideally, we would like
to generate an end product similar to that which a ‘3D’ colour-sensitive detector would
provide. The actual data storage format may (or may not) be different in detail, but
functionally equivalent. At the same time, we would like to preserve the integrity of
raw data, which, amongst other things, means avoiding degradation through excessive
Figure 2.27. Key IFU data reduction steps, in no particular order.
or inappropriate resampling. The whole process should be optimized, depending on the

project, to meet suitable signal-to-noise or resolution criteria. Finally, as well as rectify-
ing individual frames, we often need to combine exposures taken at different telescope
pointings, airmasses or instrument configurations.
Given that the first of these goals usually demands resampling, whereas the second
suggests avoiding it, one often has to prioritize one of these over the other, especially
where images are undersampled by the IFU. This is a key reason for the co-existence of
‘datacube’ and ‘Euro3D’ formats for processed data (see Section 2.6).
2.5.1 Common steps

The sequence in which data are reduced varies significantly depending on wavelength
and instrument, but there are a number of typical steps (Figure 2.27); many are common
to other spectroscopic modes, but have different details for IFS. In the remainder of this
lecture, I consider each operation in turn, roughly in the order of application. Although
details vary between instruments, the concepts are fairly generic. In the next lecture,
I shall give some examples of how the steps can be combined in a complete reduction
process.
2.5.2 Detector linearity and saturation

Optical CCD detector counts often scale almost linearly with flux until close to satu-
ration, which typically occurs when the counts reach a fixed maximum level (such as
216 − 1 = 65,535 counts), corresponding to the limit of the analogue-to-digital converter.
Any pixels read-out at the saturation limit can be flagged as bad before further reduc-
tion takes place. Clearly it is preferable to avoid saturating in the first place, however,
especially where bleeding into neighbouring pixels is likely to be an issue.
Infrared detector arrays are significantly non-linear in their response to the number
of photons accumulated. At the start of the reduction process, the measured count level
must therefore be corrected to scale linearly with flux, referring to a pre-determined
calibration curve. When IR arrays approach saturation, at their full well capacity, the
counts first become unusably non-linear and often then ‘wrap around’, dropping back
a) 5000 30 b) 30
1
4000
3000 20 .75 20
2000 .5
10 10
1000 .25
0 0
200 400 600 800 200 400 600 800
Line (pixels) Line (pixels)
Figure 2.28. Fitting and dividing out the continuum shape in a calibration lamp spectrum a),
to produce a detector pixel flat field b).
towards zero again; it may therefore be more difficult to identify highly saturated pixels
automatically than in the optical.
2.5.3 Detector bias subtraction

For CCDs, the average of several bias frames can be subtracted directly from raw data,
to correct the zero point for each pixel. If the mean bias drifts measurably with time,
one can also average or fit the bias level in the overscan region of science exposures or
flats (where applicable) and use the value to correct their overall zero point.
For IR arrays, there is usually no need to do anything regarding bias, unless the
detector controller saves ‘before’ and ‘after’ reads separately, in which case those pairs
need subtracting here.
2.5.4 Consolidating and formatting the data

By an early stage in the process, data should be packaged in an optimal format for
further reduction. If the detector is read out in separate quadrants, those might need
pasting together. Any overscan columns can be trimmed after subtraction, since they do
not contain real data. If a standard bad-pixel mask is available for the detector, one may
want to store that information with each exposure as part of a data quality array (see
Section 2.6) and/or blank out or interpolate over the corresponding pixels. A variance
array can be created from the almost-raw pixel values if one intends to propagate error
information (Section 2.6). Finally, a description of the IFU sky-to-detector mapping needs
to be stored somewhere; it is often convenient to include this as part of each image file, as
a flexible image transport system (FITS) table or in header keywords. Such information
can also be attached at this stage, if not included automatically when exposures are
read out.
2.5.5 Flat-fielding
As discussed in Section 2.3, it is necessary to correct for differences in sensitivity both
between detector pixels and across the IFU field. Depending on the instrument, these
aspects can be treated separately or together.
To create a detector flat, a set of several dispersed calibration lamp flats are usually
combined, giving a signal-to-noise ratio high enough that the flat will not contribute
random noise when applied to science data. The bias should, of course, be removed as
well. The spectral continuum shape of the co-added flat may be fitted and divided out at
each spatial position, leaving only small-scale pixel-to-pixel differences (Figure 2.28); any
larger-scale variation in detector sensitivity with wavelength is usually dealt with during
flux calibration, rather than here. The resulting flat should be normalized to an average
Figure 2.29. Detector fringing in optical fibre IFU spectra: wavelength increases to the left; the
narrow horizontal stripes are individual fibres; the thicker dark bands are gaps between blocks
of 50 fibres; a small subsection of the detector is shown.
of one, of course, to preserve flux. Raw science spectra are then divided pixel-for-pixel
by the flat, either separately or as part of a combined pixel+IFU correction (see below).
Particularly for CCDs at redder wavelengths, it may be necessary to deal with fringing
due to optical interference (e.g. Figure 2.29). This occurs when light is reflected back
and forth within the detector substrate (or perhaps between filter surfaces) before being
absorbed. For spectroscopy, this is a multiplicative effect; each point on the detector
receives quasi-monochromatic light, so fringing is a function of the total amplitude at
the relevant wavelength. In principle, fringing can therefore be removed by flat-fielding
the detector. However, the process then becomes sensitive to any small drifts in flexure
or grating tilt, since differences in illumination between science exposures and flats alter
the pattern, leaving systematic residuals.
For spectrographs that use a mosaic of separate detectors, any chip-to-chip differences
in flat-fielding must be treated carefully. Colour sensitivity curves can vary between
CCDs, for example, leaving discontinuities that cannot be corrected by scaling each chip
or fitting a smooth lamp continuum across the chips. Instead, the quantum efficiency of
each CCD may need to be factored in as a function of wavelength. Fringing properties
can also vary significantly between chips.
An IFU flat corrects the relative transmissions of different IFU elements, or variations
along image slices, to the same system (Figure 2.30). In the absence of relative flexure
between flats and science exposures, pixel-to-pixel detector variations can also be left
in, giving a combined pixel+IFU flat. This requires a calibration lamp spectrum, since
twilight flats have many sky lines and are thus unusable for calibrating spectral variations.
It is usually possible to match the flexure of science data by taking flats often enough.
This also allows IFU elements to be traced easily by using the flats as a reference. Where
there is flexure between IFU flats and science data, however, both should be divided
by a normalized detector pixel flat, removing super-Nyquist variations, before shifting
them to match. To obtain sufficient signal in between the centres of IFU elements, the
detector flat may need to be taken without the IFU in the beam, perhaps using a long
slit matching the spectral illumination of the detector by the IFU.
Where a calibration lamp does not accurately mimic the illumination of the IFU by the
sky, it can be combined with a twilight exposure to correct the spatial content. The lamp
flat provides detector pixel variations and any differences in wavelength response between
IFU elements, whilst the twilight contributes spatial variations between IFU elements,
or just a large-scale flatness correction to the spatial information from the lamp flat. The
Figure 2.30. Example IFU spectra before and after flat-fielding out detector and IFU structure.
From top to bottom, the panels are a) a single image slice, with structure from the slicing optics,
b) the same exposure after a lamp flat is applied, c) a stack of extracted fibre spectra, one per
row and d) the same data after a combined lamp and twilight flat is applied. Twilight spectra
have been used for illustration, since they are bright and highlight structure clearly.
spectral content can be removed from twilight spectra by summing in wavelength, which
also helps with signal-to-noise requirements. For consistent results, it is important to
sum over the same range of wavelength for each element, rather than a fixed number of
pixels. One can then compare the ratios of spatial variations in twilight and lamp flats,
adjusting the latter to match the former.
If IFU elements have differing transmission spectra, the lamp flat should be normal-
ized to account for this, dividing each spectrum by the average over the elements of the
continuum spectrum for the lamp. This preserves spectral corrections to the overall flux
calibration (which is derived from a standard star). It may be necessary to shift and scale
the continuum to match each element. Otherwise, if spectral differences are negligible,
the continuum can be divided out separately from each IFU element, leaving only pixel
variations. One would not normally expect to encounter large wavelength-dependent dif-
ferences across an IFU, but the possibility is worth considering, to avoid extra systematic
errors.
When reducing data from optical fibres and lenslet arrays, one can normally extract
each fibre or lens spectrum to 1D before flat-fielding. The flat can still incorporate pixel
variations, as long as science and flat-field spectra are extracted with the same weights. If
neighbouring lenses or fibres are to be deconvolved in science spectra, the same procedure
should be applied to the flat. Each spectrum is then divided through by the final flat as
normal. An exception is the unusual case of unresolved fibres, which can be treated more
like image slicer data (Turner, 2001). In principle, a possible complication with fibres is
that their relative throughputs could vary with flexure, depending on exactly how each
one is stressed (e.g. Oliveira et al., 2005). In practice, this has not been noticeable in the
author’s experience, but it gives yet another motive for taking flats to match the flexure
of science exposures.
In image slicer data, each slice can be treated much like a slit spectrum, but accounting
for differences in transmission between slices as well as along their lengths. A 2D flat
can be derived for each slice by dividing out the spectral continuum of the lamp from
each image row. To preserve spatial variations, a consistent normalization should be
maintained both along and between slices when doing this.
2.5.6 Cosmic ray and bad pixel correction
Before interpolating or co-adding data, any bad pixels should be masked out to avoid
contaminating good values (or generating artifacts). In principle, this can be done either
in 2D or 3D, with the latter providing more information at some expense in simplicity.
Any exposures that are genuinely identical (apart from noise) can simply be co-added
with statistical rejection in 2D. This often applies only for short integrations, however,
since the telescope pointing, airmass and flexure must all remain constant over at least
three frames. An alternative is to use PSF (band limitation) constraints to detect fea-
tures in individual frames that are too narrow or steep to be real (e.g. Rhoads, 2000).
This is more difficult for IFUs that pack information onto the detector with little or no
redundancy, such as image slicers, but easier for smooth and well-sampled fibre or lenslet
PSFs.
Following a similar approach in 3D, one could identify bad pixels in the process
of constructing a datacube. This makes a greater number of neighbours available for
constraining (and interpolating) values, but does not allow for interpolation prior to
3D reconstruction. When co-adding data in 3D, statistical rejection can be applied
to the interpolated values and used alongside data quality arrays from an earlier stage,
to identify bad samples.
Given the small number of spatial elements and high density of information in a typical
IFU dataset, defective pixels may have a relatively high impact. At the same time, they
can be less straightforward to identify than when co-adding direct images, extracting a
2D spectrum of a point source or working with more extended slit spectra. This reduction
stage may therefore require some care.
2.5.7 Instrumental background
To ensure that measurements reflect the true object+sky intensity, any instrumental
background counts must be removed. Especially with fibres, nearby elements at the IFU
input can be widely separated at the detector, so it is possible for varying background
levels to introduce structure on small spatial scales. Where there are gaps between IFU
elements (or blocks of elements) at the output, the background level in the gaps can be
used to model and remove any diffuse scattered light. In the infrared, however, dark cur-
rent and any other target-independent background sources are normally taken care of by
sky subtraction, without special treatment. Note that background light can have differ-
ent colour characteristics from the dispersed object signal and change between mosaiced
detectors.
2.5.8 Spatial calibration
Before data can be transformed into a physical coordinate system, it is necessary to
identify IFU elements (fibres, lenses or slices) on the detector and calibrate their spatial
coordinates. This usually involves locating the spatial elements in a flat-field spectrum
matching the science data, with reference to a known IFU mapping.
For individually resolved optical fibres, one can identify fibre peaks in a cross-section
perpendicular to the dispersion axis (see Figure 2.7) and trace them along the length of
the detector. If individual spectra are heavily blended together, one can use any gaps
between blocks of fibres or perhaps cross-correlation of the fibre throughput pattern to
calculate where individual elements are located. If there is no flexure (e.g. at a Nasmyth
focus), the positions can be pre-determined by masking out groups of fibres in a flat and
tracing one group at a time.
For lenslet IFUs, we are again looking for peaks in cross-section, but with the spectra
distributed around the detector in 2D rather than forming a slit. An optical model is
typically used to determine the relative positions and lengths of spectra for a given sky-
to-detector mapping, reducing the problem to one of measuring zero-point offsets from
the peaks.
For image slicers, any gaps between the slices at the output can be used to find their
edges. A special Ronchi screen (consisting of regular parallel lines) may also be placed in
front of the IFU input, to calibrate the alignment and scale of each slice along its length.
If slices are staggered in wavelength, steps in the wavelength solution could provide an
alternative indication as to where the edges are. In the absence of flexure, pre-determined
pixel positions can again be used. Having found the edges of the slices (or Ronchi bands),
the tilt and curvature of each one can be traced along the dispersion axis of the detector.
The usable length of each slice can be delimited according to some data threshold or a
fixed number of pixels from the centre.
2.5.9 Wavelength calibration

Depending on the type of IFU, a set of 1D or 2D spectra have to be calibrated in
wavelength. Fibre spectra can usually be calibrated individually after extraction to 1D.
Lenslet spectra are sometimes calibrated to first order as part of the extraction process;
the results can be refined later in 1D if needed. For image slicers, we can fit coordinates
separately for each 2D spectrum. Whatever the format, the wavelength reference is most
often an arc lamp spectrum (Figure 2.8), processed in the same way as science data.
This is used to fit a relatively low-order curve in 1D or 2D for each spatial element,
describing the wavelength as a function of pixel index. The procedure is similar to that
for multi-object spectroscopy with fibres or slits, except that manual interaction may
not always be practical for IFUs with thousands of spectra to calibrate, in which case
reliable automatic line identification is more critical.
For science data containing identifiable sky lines, these can be used to account for
any flexure with respect to the arc exposures, applying a small zero-point correction to
the full wavelength solution from the arc (otherwise, arcs must be taken often enough
to match science data). At infrared wavelengths, there are often enough sky lines over
the length of a spectrum to fit a full wavelength solution without having to use an arc
at all. Once a wavelength has been determined for each raw or extracted sample point,
there are two options for rectification (see also Section 2.5.10). Typically, individual 1D
or 2D spectra are transformed to a common, linear grid before continuing with further
processing. It is also possible, however, to map the points directly to 3D without an
intermediate resampling stage.
Given the presence of optical distortions across the spectrograph and/or step offsets
between image slices in some designs, it is quite common for different IFU elements to
have slightly different spectral ranges. This leads to a dataset covering only parts of the
IFU field at the ends of the spectra.
2.5.10 Extraction
This stage involves extracting IFU elements from the raw data format to something
convenient for later processing. For fibres or lenses, we usually want to extract each
spatial element to a 1D spectrum, integrating in cross-section with or without weights
(Figure 2.31). This is complicated by the fact that adjacent spectra often overlap substan-
tially in cross-section (e.g. Figure 2.7). For fibres aligned in wavelength, this amounts to
an extra convolution in one spatial dimension: the simplest option is to accept a small loss
of spatial resolution, summing over each fibre along with some signal from its neighbours.
Where spatial resolution is exceptionally important, the alternative is to try deconvolv-
ing neighbouring spectra, at a significant cost in extra noise. Deconvolution is sensitive
Figure 2.31. Example of GMOS fibre spectra before and after extraction to 1D by summing
over each fibre in cross-section.
to knowing the PSF accurately, which is complicated by variations between fibres or

lenslets, but has been used successfully for IFUs such as CIRPASS and OASIS. The
unusual case of very closely spaced (∼ 1 detector pixel) fibres can be treated more like
an image slicer, without extracting to 1D (Turner, 2001).
Since adjacent fibres can have different throughputs, spectra extracted without decon-
volution receive slightly different contributions from the fibres on either side. This per-
turbs the effective spatial coordinate of each fibre slightly along one axis, producing a
sampling grid that is not quite regular on the sky. In practice, this effect may be very
small, except close to dead fibres, and is sometimes safe to ignore.
For lenslet IFUs (or misaligned fibres), any overlaps between spectra are offset in the
dispersion direction. Rather than just causing slight spatial smoothing, this contaminates
spectra with light from the wrong wavelengths. It is therefore important to deconvolve
any signal from neighbouring lenslets in this scenario, to avoid spurious measurements.
For image slicers, extraction generally does not involve summing over pixels, but sooner
or later one wants to ‘cut out’ the useful regions of raw images corresponding to the slices.
This is done after tracing the slice positions, before or after wavelength calibration. One
approach is to straighten each slice, producing a traditional 2D extracted spectrum, as
for a slit. This may be convenient for further processing with existing software. All the
slices should be resampled onto the same grid in physical (e.g. x, λ) coordinates, so they
can simply be stacked later in 3D along the other spatial axis (y). Even so, a further
resampling stage may still be necessary, to correct atmospheric dispersion or combine
datasets from different pointings. The other possibility for ‘extraction’ is to map pixel
values directly from the original raw format to a 3D representation, after calibrating in
wavelength, without any intermediate resampling in 2D. The latter method is appropriate
for minimizing the interpolation of undersampled data.
Figure 2.32. Section of a near-infrared image slicer spectrum, before and after sky subtraction.
Many bright pixels with high dark current also subtract out well.
2.5.11 Sky subtraction

To measure the signal from a science target, sky emission lines and continuum must
first be removed. For infrared (or optical ‘nod-and-shuffle’) data, sky and any other
background are normally removed by subtracting pairs of consecutive raw exposures early
on (e.g. Figure 2.32). This means that corresponding object and sky measurements are
taken with the same IFU elements and detector pixels, avoiding any subtraction residuals
due to flat-fielding errors or slight mismatches in spectral profiles between different IFU
elements. If the sky varies significantly between exposures, one can attempt to interpolate
between sky frames taken before and after the target object, or if darks are available,
it may be feasible to scale the sky to the correct level. Telluric line species can vary
differently, however, so may not always scale well together.
For most optical data, sky is subtracted from 1D/2D calibrated spectra or a 3D data-
cube. The sky spectrum can be averaged over a number of spatial pixels with no object
signal, to reduce the noise level, before subtracting it from all the spatial pixels. Thus
the process adds little extra noise to the result. Depending on the instrument and target,
the sky spectrum can be taken from separate sky fibres, the edges of the science field
or a blank sky integration. Since different optical fibres or lenslets can have slightly
different PSFs, subtraction residuals may be an issue (Allington-Smith and Content,
1998). Convolving object and sky spectra to the same resolution could help avoid this,
at a cost. One could also attempt to match up IFU elements with similar PSFs, either
empirically or nearby on the detector, where optical variations within the spectrograph
are an issue. This has not yet been tested, to the author’s knowledge.
2.5.12 Telluric correction

In red or infrared spectra, any telluric absorption lines that obscure the true target
spectrum need removing. To obtain a 1D telluric reference spectrum, a standard star
observation is reduced in the same way as scientific data and its flux is summed up spa-
tially. If necessary, one can try to match the airmass of a science exposure by interpolating
between telluric spectra taken at different airmasses. The spectrum can be normalized
by fitting and dividing out the continuum shape (interpolating over any weak stellar
absorption lines). The result is then used to divide out telluric features from 2D cali-
brated science spectra or a 3D datacube (Figure 2.33). If the spectral PSF varies over
Figure 2.33. Sketch illustrating correction for telluric absorption.
the IFU, one could dither the standard around the field and try extracting multiple
spectra, for calibrating different groups of elements.
2.5.13 Flux calibration

Here, a standard star is used to correct measured counts for the combined transmis-
sion of the instrument, telescope and atmosphere as a function of wavelength. As for
telluric calibration, we reduce a standard star observation with the same instrument
configuration as the corresponding scientific data and sum up flux spatially, to get a
1D spectrum. The same dataset might, in fact, be used for both purposes. To derive a
sensitivity or throughput spectrum (e.g. Figure 2.10), one must somehow determine the
intrinsic flux of the standard star as a function of wavelength and compare this with
the measured counts. In the optical, one can refer to tabulated narrow-band fluxes that
are available for numerous stars in the literature (and in standard reduction software).
In the infrared, one can use a published broad-band magnitude and spectral type (i.e.
temperature) to generate a model black-body curve matching the continuum of the stan-
dard. Reduced scientific data in 1D–3D are then divided through at each spatial point by
the derived sensitivity function, converting measured counts into fluxes on a consistent
system such as W m−2 Å−1 arcsec−2 . Where either standard star or scientific observa-
tions have been taken through cloud, normalizing to some arbitrary scale might be more
appropriate. Although much variation of the sensitivity spectrum over the IFU field is
unlikely, appropriate flat fielding can take care of any differences (Section 2.5.5).
2.5.14 Reconstruction in 3D
Reduced data in 1D or 2D arrays can be transformed back to a physical coordinate grid
either before or after scientific analysis, depending on the application. In general, this
requires interpolation. One approach is to resample once onto a 3D datacube in (x, y, λ),
for subsequent analysis and visualization. Special tools (such as the Euro3D software)
can also be used to interpolate ‘on the fly’, as needed during visualization or analysis.
Where applicable, a third possibility is to analyse individual 1D spectra and interpolate
the results onto a 2D map. Finally, one may prefer to perform analysis directly in 3D
at the original, calibrated pixel positions, perhaps by fitting a physical model of the
target. This last method avoids resampling and any associated uncertainties, but makes
it difficult to use existing, general-purpose software.
The best process for transforming data to a uniform grid depends on both the initial
sampling and IFU geometry. For well-sampled datasets, one can interpolate the data
with smooth surfaces that reconstruct the dispersed telescope image accurately, allowing
Figure 2.34. a) Example of extracted fibre IFU data from a Seyfert galaxy observation and
b) one image plane of the datacube produced by resampling spatially with smooth, piecewise
polynomials.
resampling onto datacubes or 2D spatial maps with minimal degradation (Figure 2.34).
The result is therefore similar to what a 3D detector might provide more directly. For
undersampled data, extra care must be taken to avoid introducing aliasing artifacts or
excessive smoothing; it is not possible to reconstruct exactly what a 3D detector would
see. One should resample as little as possible – maybe once, to get to 3D – and stick to
positive, compact interpolants, such as nearest-neighbour, linear or Drizzle weightings.
Depending on the geometry of the IFU, one has a choice as to how many dimensions
to interpolate in at once. Resolved fibre or lenslet spectra can be extracted to 1D and
resampled onto a common grid in wavelength before interpolating spatially. Where the
IFU input consists of square lenslets, the spectra can then simply be arranged to form a
2D grid spatially. For lenslets or fibres packed in a triangular (hexagonal) configuration,
a square spatial grid is obtained by interpolating between the spectra at each wavelength
element, in either 1D or 2D. To deal with one dimension at a time, it is possible to
interpolate first along each linear row of hexagons or circles and then across the rows.
With image slices, the (x, λ) coordinates of each 2D spectrum are not trivially separa-
ble, due to image distortions. It therefore makes more sense to interpolate spatially and in
wavelength at the same time. This is the same problem as straightening and wavelength
calibrating a long-slit spectrum in 2D, except that all the slices should be transformed
to the same (x, λ) grid, so cannot be treated completely independently. Once the slices
are rectified, they need just to be stacked in y to get to 3D. Image slicers often have
rectangular spatial pixels, however, with the slice width being different from the detector
pixel size. In those cases, one can interpolate between slices in y, to get to a square spatial
grid.
For any type of IFU data, the alternative to resampling each element individually is
to interpolate processed detector pixel values directly in 3D, after calibrating x, y and λ
for each pixel. This is typically more complicated (except for the simplest interpolants)
than working in lower dimensions, especially where samples are distributed irregularly
in physical coordinates. For interpolating smooth curves, it is significantly easier to find
pre-existing software that works in fewer than three dimensions, on regular grids and
with homogeneous coordinates. For undersampled data, however, transforming straight
to 3D may be the best approach, using something like a nearest-neighbour or Drizzle
scheme to map pixels directly onto a cubic grid. As of 2006, the Gemini IRAF (image
reduction and analysis facility) package includes a task, ‘gemcube’, that can do this using
a version of the Drizzle algorithm in 3D.
2.5.15 Atmospheric dispersion
Atmospheric dispersion, or differential refraction, causes the position of a target within
the IFU field to vary with wavelength. This is typically less problematic than for a single
slit, where the target can be shifted entirely out of the aperture at some wavelengths.
Nevertheless, the effect can distort spectra and cause different-sized variations at different
airmasses. Where significant, a correction must therefore be applied before combining
datasets from different airmasses or comparing the positions of spatial features at different
wavelengths.
The spatial offset as a function of wavelength can be determined either using an atmo-
spheric refraction model (given the airmass, position angle and other parameters) or by
centroiding on a point source within the IFU field at different wavelengths. Some details
are discussed in Arribas (1999). After constructing a datacube, different corrective shifts
can then be applied to each wavelength plane. Alternatively, the sky coordinates of each
pixel (as calculated during spatial calibration) can be recalculated before transforming to
3D or comparing with a model. At low airmass or infrared wavelengths without adaptive
optics, it may be safe to assume dispersion is negligible. Also, when integrating flux over
a spatial aperture centred on a point source, dispersion is taken care of without special
treatment.
2.5.16 Dithering and mosaicking

Dithering and mosaicking are discussed in the accompanying lectures by Pierre Ferruit,
so are mentioned here only briefly. This stage often involves shifting and adding already-
resampled data in 3D, but can also be integrated with 3D reconstruction, resampling
multiple exposures onto the same grid. The latter approach would be better with spatial
undersampling and/or for combining sub-pixel dithers. Occasionally, where exposures are
taken with integer-spatial-pixel offsets and negligible differences in flexure or airmass, it
may be possible to co-add extracted spectra instead of 3D data.
2.5.17 Spatial binning

For IFUs with small spatial pixels, it is often difficult to obtain a good signal-to-noise
ratio across the full field. Targets such as galaxies can vary spatially over a large range
of brightness within a few arcseconds. To make optimal use of the available information,
one may want to bin data adaptively, in variably sized spatial regions. The Voronoi
tessellation method of Cappellari and Copin (2003) divides IFU data into irregular bins of
approximately constant minimum signal-to-noise ratio, allowing reliable spectral mapping
over the whole field (Figure 2.35). This sacrifices critical sampling of the seeing, of course,
so it is important to interpret spatial content accordingly, perhaps by comparison with
a similarly binned model of the target. If the object is known to be intrinsically smooth,
this may be less of a concern.
2.5.18 Summary of data reduction issues

A number of reduction stages are the same as, or similar to, those for other spectroscopic
modes. IFUs have a few rather special requirements, however, notably for flat-fielding and
mapping data from 2D to 3D. The details of how sky subtraction is done in the optical
and how spectra are located and extracted are also somewhat different. Dithering and
mosaicking steps are more akin to imaging reduction than slit spectroscopy. Lastly, unlike
conventional imaging or spectroscopy, IFS allows (and sometimes requires) after-the-fact
correction of atmospheric dispersion.
20
10
Arcsec
0
−10
−20
−20 −10 0 10 20
Arcsec
Figure 2.35. A spatial Voronoi tessellation of IFU data, using bins of approximately constant
signal-to-noise ratio (Cappellari and Copin, 2003).
As a final note, it is worth emphasizing that, with the correct treatment, one can often
use IFS data for high-fidelity imaging, as well as spectroscopy.
2.6 Data reduction process

The last lecture gave an overview of the different stages of basic IFU data reduction. Here,
I discuss the data reduction process as a whole, showing some brief example sequences
for optical and infrared data and considering some associated practicalities, namely error
and data quality propagation and file formats.
2.6.1 Reduction sequence

Figure 2.36 and Figure 2.37 show typical optical and near-infrared data reduction
sequences, respectively, for two of the Gemini IFUs. Data from other well-sampled opti-
cal fibre or infrared image slicer IFUs could be reduced similarly. Separate sequences for
calibration data such as flats, biases and standard stars are not shown here, but these
usually follow the same process part way through, before or after combining exposures.
For example, optical flats are reduced up to the stage of bias subtraction and co-added.
Standard stars are usually treated like scientific exposures, except for flux or telluric
correction, before extracting a spectrum from the 2D or 3D dataset.
2.6.2 Error propagation

At the end of a data reduction sequence, it is often important to have a good estimate
of the errors in data values. This allows us, for example, to estimate the statistical
significance of a faint detection, the reliability of stellar ages and metallicities derived
from line strength indices (Cardiel et al., 1998) or the intrinsic random errors in velocity
measurements.
Raw data have relatively well-defined errors, due to photon statistics and detector
read noise – but after numerous processing stages it is at best difficult and at worst
impossible to estimate errors directly from the data values. The solution is to track
the errors throughout processing. For each detector pixel, an error value is stored in a
separate image, alongside the main data array. During each processing step, this array
Figure 2.36. Example reduction sequence for scientific exposures from the GMOS optical fibre
IFU, using Gemini IRAF package tasks.
is processed in parallel with the scientific image, to reflect how the errors change. When
adding two images, for example, the corresponding errors are added in quadrature.
Sources of error
A process where discrete values vary statistically around a well-defined mean, such as
counting photons, is described by a Poisson distribution:
µn
Pµ (n) = e−µ , (2.12)
n!
Figure 2.37. Example reduction sequence for scientific exposures from the GNIRS infrared
image slicer, using Gemini IRAF package tasks. Telluric and flux correction are not shown, but
in this case would be performed between nstransform and nfcube.
with a mean (expected number of photons) of n = µ. The standard deviation from the
√
mean is simply σ ≡ µ, so when counting photons (really photo-electrons), the statistical
error is the square root of the expected number of photons (electrons). In practice, the
error is estimated using the square root of the measured number of electrons, since that
√
is what we know. For large µ, the Poisson distribution is Gaussian, with σ = µ.
The main random sources of measurement error in data are read noise from the detector
electronics, Poisson noise from the science target and sky (or other background) and
Poisson noise from detector dark current. The last two of these can be estimated as
discussed above, directly from the pixel values, whilst read noise is a fixed property of
the instrument. We also have systematic errors – notably inaccuracies in flat-fielding –
introduced during processing or by instrumental variations. These are usually present at
the level of a few percent and are difficult to reduce to zero. Such effects can be more
difficult to account for than noise, but are typically dominated by statistical errors.
Creating and propagating error arrays

Detector electronics usually do not report one count per stored electron. The number of
electrons corresponding to a single count, or ‘analogue-to-digital unit’ (ADU), is known
as the detector gain (really an inverse gain). This is set to some value low enough for read
noise to dominate digitization errors, but high enough to avoid reaching the maximum
value of the analogue-to-digital converter at too small a fraction of the detector full-well
capacity. Typical gains vary from just a few e− /ADU for CCDs to > 10 for IR arrays,
the latter having greater read noise and higher backgrounds to deal with. Since Poisson
statistics apply to electrons, rather than ADUs, measured counts must be multiplied by
the gain before taking the square root to estimate Poisson noise (in electrons).
Given that errors add in quadrature when adding or subtracting data, the numbers
stored in an error array are actually variance (σ 2 ) values, rather than the noise (σ) itself,
to avoid repeated square and root operations. If science data are stored in ADUs, the
corresponding variance arrays must also be converted to those units. Alternatively, science
data can be multiplied through by the gain at the start of reduction and everything
propagated in units of electrons.
As a starting point for error propagation, a variance array can be created containing
2
the square of the detector read noise, σread , which affects every pixel independently of its
count level. Read noise is measured in electrons, so when working in detector counts, the
initial variance is then (σread /g)2 , where g is the gain. After subtracting the electronic
bias level from raw data, one can add the statistical variance in the measured counts, n,
for each pixel, which is ng in electrons or n/g in ADUs. This sum represents the total
variance for the raw data, which is therefore (in ADUs)
2
σread ntotal − nbias
σ2 = + . (2.13)
g2 g
At each subsequent reduction step, the variance array is manipulated according to the
operation being performed on the main science array. When adding or subtracting pixels,
their errors add in quadrature:
σ 2 = σ12 + σ22 . (2.14)
When scaling an image by some factor, a, the error is scaled accordingly:
σ 2 = a2 σ12 . (2.15)
When one image is multiplied or divided by another, the fractional errors add in quadra-
ture. Each variance array must therefore be divided by the square of the corresponding
science image, adding the results and multiplying the sum by the square of the final
science image:
() * ) *2 +
2
2 2 σ1 σ2
σ =n + . (2.16)
n1 n2
For more complex operations on science data, represented by some arbitrary function,
n = f (n1 ), the first derivative of f can be used to estimate how output values vary with
small changes in the input for a given n:
σ 2 = f ′ (n1 )2 σ12 . (2.17)
At the end of the complete reduction process, error propagation allows the final noise
values to be determined relatively accurately just by taking the square root of the final
variance array.
Resampling
In raw data, each pixel has an independent statistical error. If resampling causes smooth-
ing, however, the errors of different pixels become correlated. As an example, linear inter-
polation at the mid-point between samples √ is an averaging of pairs. The error of each
interpolated value is therefore reduced by 2, without halving the number of pixels.
Clearly the average signal-to-noise ratio over all pixels cannot have increased, however,
nor
√ can summing two adjacent resampled pixels reduce the error by a further factor of
2, since only three of the original values are involved. What has happened is that the
errors of adjacent pixels are no longer independent.
In a case like this, one could track the errors by propagating a higher-dimensional
covariance matrix, where the covariance of two variables is defined as the expected value
of the product of deviations from the means:
σxy = ⟨(x − x̄)(y − ȳ)⟩ = ⟨xy⟩ − x̄ȳ. (2.18)
The variance of a single variable is therefore equivalent to its own covariance:
σx2 = σxx = ⟨x2 ⟩ − ⟨x⟩2 . (2.19)
,
With this information, the error on a sum ni=1 xi is calculated using
n
$ $
σ2 = σx2 i + 2 σx i x j . (2.20)
i=1 i< j
For more information, see an error analysis or statistics textbook, such as Taylor (1997).
Since this level of error propagation is computationally complex, reduction software
usually does not track covariance. It is nevertheless important to be aware that simple
variance estimates may not be exactly correct after resampling. In the ideal limit of
resampling a sinc interpolant at the original rate, however, covariance is not an issue
because no smoothing occurs (at any given point, the RMS distance of positively and
negatively weighted sample contributions is zero).
2.6.3 Data quality

As well as storing variance data alongside each science image, it is useful to include
‘data quality’ information. A separate integer array is used to track which pixels are
good or bad (and why) in the main science array. Each bit of the integer represents a
Boolean yes/no flag for a particular defect, allowing for multiple problems with each
pixel. For example, a value of zero typically represents a good pixel, with different bits
set to one where issues such as cosmic ray contamination, saturation or high dark current
are identified. Clearly, some of these result in total loss of information in a pixel, whereas
others just lead to noisy values.
Data quality propagation can help significantly with masking out values at each reduc-
tion stage, given that defective samples may be impossible to identify directly from their
values after processing steps such as interpolation or sky subtraction. Perhaps unfortu-
nately, the convention as to which bits represent what defects depends on the software.
2.6.4 File storage format

Like other astronomical data, IFU spectra are typically stored and processed in ‘flexible
image transport system’ (FITS) files. This standard is overseen by an IAU working group
and documented at http://fits.gsfc.nasa.gov. Each single FITS file can contain multiple
datasets, along with information describing the data. The most common elements are
! one or more N-dimensional image arrays;

the following:
! ASCII header information, using keyword=value pairs of multiple data types
! one or more binary tables, using named columns (e.g. XCOORD) and mixed types
(e.g. OBJECT = ‘NGC1068’ or EXPTIME = 120);
to store heterogeneous data.

EXT# EXTTYPE EXTNAME EXTVER DIMENS BITPIX OBJECT

0 NGC1068
1 BINTABLE MDF 32x21 8
2 IMAGE SCI 1 32x1022 -32 NGC1068
3 IMAGE VAR 1 32x1022 -32 Variance
4 IMAGE DQ 1 32x1022 32 DQ
5 IMAGE SCI 2 32x1022 -32 NGC1068
6 IMAGE VAR 2 32x1022 -32 Variance
7 IMAGE DQ 2 32x1022 32 DQ
[ ... etc ... ]
Figure 2.38. Example multi-extension FITS structure for part-processed image slicer data,
from the edited output of the IRAF task fxheader. The full file continues until EXTVER=21,
corresponding to the 21st image slice. The ‘MDF’ (mask definition file) extension is a binary
table containing the IFU sky-to-detector mapping. Fibre or lenslet data are more likely to be
stored in a 2D or 3D array in a single EXTVER.
Multiple datasets within a FITS file are kept in separate ‘extensions’. First is a primary
header, containing keywords applicable to the whole file, such as the object name, tele-
scope pointing, airmass, filter and central wavelength. This is followed by one or more
numbered and/or named extensions, each containing an image array or binary table,
along with any header keywords specific to that particular dataset. Figure 2.38 shows an
example FITS structure for image slicer data, including variance and data quality arrays
and a table with details of the IFU mapping.
2.6.5 Formats for reduced data

Inside a FITS file or its equivalent, fully reduced IFU data can be stored in one of at
least three formats for subsequent analysis. A basic option for extracted 1D spectra is
to work with ‘row-stacked spectra’, each of which is stored as a separate row of a single
2D image. Without further image reconstruction, this method is limited to 1D spectral
analysis such as velocity mapping. One then has to create a 2D spatial map (such as a
velocity field) later on from the results.
An IFU datacube is a 3D image with two spatial axes and one wavelength axis, usually
arranged in that order. It is therefore a natural extension of 2D image storage. As a simple
array, the format is straightforward to read and manipulate in IRAF, Python, IDL or
any other generic data processing environment. Generating the datacube usually requires
resampling processed IFU data onto a 3D grid (unless the IFU has an already square
lenslet array). It is therefore possible that a datacube could represent the original data
less faithfully than a minimally processed format, due to interpolation-related artifacts
or smoothing. On the other hand, it provides an opportunity for resampling to be done
carefully just once, at a well-defined stage. Where interpolated data are oversampled to
produce ‘smooth’ images for visualization, datacube file sizes can become quite large,
which can lead to redundant processing later on as well as requiring more storage than
interpolating ‘on the fly’. By resampling close to the Nyquist rate, however, datacubes
can be made at least as compact as other formats.
The Euro3D data format was developed as a standard for exchanging data between
reduction and analysis software packages, without requiring spatial resampling and using
a relatively sophisticated storage arrangement. This is the native format of the ‘E3D’
D A
T
T
A
A
Q
ST
D
(Potential) vector
Example of valid data
ID
Sp #
C o d1
G 2
el
up
d
or
or
ax
ro
Co
D
ti h
le c I
ex
ar gt
Se trum
NS tion
St le n
nd
Sp X
PA
ec
ec
Sp
Figure 2.39. The table structure of the Euro3D data format (From Kissler-Patig et al., 2003).
visualization tool (although the latter can also read datacubes). Both 1D or 2D spectra
and the attributes describing them are stored together in a binary table (Figure 2.39).
The format is closer to the raw data than a cube, avoiding resampling until needed during
visualization or analysis. The file size is therefore minimal, as for row-stacked spectra.
One limitation is that special libraries are needed to work with the format, although
these are now publicly available in the C programming language from the Centre de
Recherche Astronomique de Lyon.
2.6.6 Summary of the data reduction process

The reduction sequences shown here for optical and near-infrared IFU data should give
an idea how to tie together the steps discussed in Section 2.5. For more specific details,
it is best to refer to reduction examples or software documentation for the instrument
being used, which are tailored to that instrument’s particular characteristics.
Most data reduction software packages operate on FITS files, each of which has one or
more image extensions. Software packages sometimes have the option to propagate error
and data quality arrays alongside each image, to help understand the accuracy of results
better and perhaps also optimize data combination. The final output of the process,
ready for analysis, is usually a set of row-stacked spectra, a datacube or a Euro3D file.
The best option amongst these depends on the analysis software, scientific application
and user preference. For those working in the Euro3D framework, the data format speci-
fication, libraries and visualization tool are now available under the appropriate links at
http://www.aip.de/Euro3D.
2.7 Acknowledgements
Thanks to the organizers for their kind invitation to the Winter School. My initial studies
of IFU data reduction and sampling were aided by input from Jeremy Allington-Smith,
Robert Content and the instrumentation group at Durham, UK. The Gemini Observatory
has since given me the opportunity to commission and observe with numerous IFUs. Some
of the details in these lectures are based on my work or discussions with Bryan Miller at
Gemini, Frank Valdes at NOAO, the Euro3D team and other colleagues.
This work was supported by the Gemini Observatory, which is operated by the Associa-
tion of Universities for Research in Astronomy, Inc., on behalf of the international Gemini
partnership of Argentina, Australia, Brazil, Canada, Chile, the UK and the USA.
REFERENCES
Allington-Smith, J.R., Content, R. (1998), PASP, 110, 1216
Allington-Smith, J.R., Murray, G., Content, R. ET AL. (2002), PASP, 114, 892
Allington-Smith, J.R., Dubbeldam, C.M., Content, R. ET AL. (2007), MNRAS, 376, 785
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B. ET AL. (1999), A&AS, 136, 189
Bershady, M.A., Andersen, D.R., Harker, J. ET AL. (2004), PASP, 116, 565
Beutler, F.J. (1966), SIAMR, 8, 328
Bowen, I.S. (1938), ApJ, 88, 113
Bracewell, R.N. (2000), The Fourier Transform and Its Applications, McGraw-Hill, 3
Cardiel, N., Gorgas, J., Cenarro, J., Gonzalez, J.J. (1998), A&AS, 127, 597
Fruchter, A.S., Hook, R.N. (2002), PASP, 114, 144
Goodman, J.W. (1968), Introduction to Fourier Optics, McGraw-Hill
Hecht, E. (1987), Optics, Addison-Wesley, 2
Kissler-Patig, M., Copin, Y., Ferruit, P., Pécontal-Rousset, A., Roth, M.M. (2003),
Euro3D data format, format definition, http://www.aip.de/Euro3D
Kotel’nikov, V.A. (1933), On the transmission capacity of ‘ether’ and wire in electrocommu-
nications, Izd. Red. Upr. Svgazi RKAA
Lauer, T.R. (1999), PASP, 111, 227
Lehmann, T.M., Gönner, C., Spitzer, K. (1999), Survey: interpolation methods in medical
image processing, IEEE Trans. Medical Imaging, 18, 1049
Marvasti, F.A. ed. (2001), Nonuniform sampling: theory and practice, Kluwer Academic/
Plenum
Moffat, A.F.J. (1969), A&A, 3, 455
Nyquist, H. (1928), Certain topics in telegraph transmission theory, AIEE Trans, 47, 617
Oke, J.B. (1990), AJ, 99, 1621
Oliveira, A.C., de Oliveira, L.S., dos Santos, J.B. (2005), MNRAS, 356, 1079
Pasquini, L., Avila, G., Blecha, A. ET AL. (2002), Installation and commissioning of
FLAMES, the VLT Multifibre Facility, Msngr, 110, 1
Raab, W., Poglitsch, A., Looney, L.W. (2004), FIFI LS: the far-infrared integral field
spectrometer for SOFIA, SPIE, 5492, 1074
Rhoads, J.E. (2000), PASP, 112, 703
Shannon, C.E. (1949), Communication in the presence of noise, IRE Proc., 37, 10
Strohmer, T. (1993), Efficient methods for digital signal and image reconstruction from nonuni-
form samples, PhD, University of Vienna
Taylor, J.R. (1997), An introduction to error analysis, the study of uncertainties in physical
measurements, University Science Books, 2
Turner, J.E.H. (2001), Astronomy with integral field spectroscopy: observation, data analysis
and results, PhD, University of Durham, UK
Vanderriest, C.A. (1980), PASP, 92, 858
Whittaker, E.T. (1915), On the functions which are represented by the expansions of the
interpolation theory, Proceedings of the Royal Society Edinburgh, Section A, 35, 181
Wright, G.S., Rieke, G.H., Colina, L. ET AL. (2004), The JWST MIRI instrument concept,
SPIE, 5487, 653
3. 3D spectroscopic instrumentation
MATTHEW A. BERSHADY
In this chapter we review the challenges of, and opportunities for, 3D spectroscopy and
how these have led to new and different approaches to sampling astronomical infor-
mation. We describe and categorize existing instruments on 4 m and 10 m telescopes.
Our primary focus is on grating-dispersed spectrographs. We discuss how to optimize
dispersive elements, such as VPH gratings, to achieve adequate spectral resolution,
high throughput, and efficient data packing to maximize spatial sampling for 3D spec-
troscopy. We review and compare the various coupling methods that make these spec-
trographs ‘3D’, including fibres, lenslets, slicers, and filtered multi slits. We also describe
Fabry–Perot (FP) and spatial-heterodyne interferometers, pointing out their advantages
as field-widened systems relative to conventional, grating-dispersed spectrographs. We
explore the parameter space all these instruments sample, highlighting regimes open
for exploitation. Present instruments provide a foil for future development. We give an
overview of plans for such future instruments on today’s large telescopes, in space and in
the coming era of extremely large telescopes. Currently-planned instruments open new
domains but also leave significant areas of parameter space vacant, beckoning further
development.
3.1 Fundamental challenges and considerations

3.1.1 The detector limit I: six into two dimensions
Astronomical data exist within a six-dimensional hypercube sampling two spatial dimen-
sions, one spectral dimension, one temporal dimension, and two polarizations. In contrast,
high-efficiency, panoramic digital detectors today are only two-dimensional (with some
limited exceptions). The instrument builder’s trick is to down-select the critical obser-
vational dimensions relevant to addressing a well-motivated subset of science problems.
Here we consider the application to 3D spectroscopy at high photon count rates, where
both spatial and spectral domains must be parsed on to, for example, a charge coupled
device (CCD) detector, as illustrated in Figure 3.1. The choice is in how the datacube is
sliced along orthogonal dimensions, since it is not easy to rotate a slice within the cube.
Such ‘rotation’ could be accomplished via multi-fibre or multi-slicer feeds to multiple
spectrographs, but to date the science motivation has not led to such a design. In prac-
tice, then, we have the extremes of single-object, cross-dispersed echelle spectrographs, to
FP monochromators. The ‘traditional’ integral-field spectrograph (IFS) is between these
two limiting domains.
In addition to balancing the trade-offs between spatial versus spectral information,
there is also the issue of balancing sampling (i.e. resolution) versus coverage in either of
these dimensions. Science-driven trades formulate any specific instrument design. When
sampling spatial and spectral domains, not all data have equal information content. Hence
one may also consider integral versus sparse sampling. Fibre-fed IFS such as Hexaflex
(Arribas et al., 1991) and SparsePak (Bershady et al., 2004) are examples of sparse sam-
pling in the spatial domain. Multi-exposure FP observations, multi-beam spectrographs,
or notch gratings (discussed below) are examples of instruments with the capability of
sparse sampling in the spectral domain.
87
88 Matthew A. Bershady
θ Slicing the datacube

λ
θ
XD Long slit Integral field Fabry–Perot

Grating, FTS, or SHS (surface is actually
aperture curved)
Integral Sparse spatial Sparse spectral

Resolution versus Fibres, lenslets sampling sampling
Slicers, interferometers Fibres or MOS Multiple exposures with FP,
Coverage (fibres or slicers) multi-beam spectrograph or
notch gratings
Tilted sampling?
Figure 3.1. Sampling the datacube with equal volumes and detector elements.
3.1.2 Merit functions

There are a number of generic merit functions found in the instrumentation literature,
in a variety of guises used, or tailored, to suit the need of comparing or contrasting the
niche of specific instruments. Some useful preliminary definitions (used throughout this
! the spectral resolution, R = λ/dλ;

chapter) are:
! the number of spectral resolution elements, N ;

! spectral coverage = ∆λ = N × dλ;
R
! spatial resolution dΩ, i.e. the sampling element on the sky (fibre, lenslet, slicer slitlet,
R
! number of spatial resolution elements, N ; and

or seeing disc);
! spatial coverage Ω = N × dΩ.

Ω
Ω
With these definitions, the trade-offs discussed above may be summarized by stat-
ing that NR × NΩ must be roughly constant for a given detector. Another important
statement is that A × Ω, or grasp, is conserved in an optical system (A is the telescope
collecting area): The same instrument has the same A × Ω on any diameter telescope
with the same focal ratio, something derived from the identify Ω = a/f 2 , where a is the
instrument focal area and f the focal length. What changes with aperture, of course, is
the angular sampling. For sufficiently extended sources, angular sampling is not neces-
sarily at a premium. Imagine, for example, dissecting nearby galaxies with a MUSE-like
instrument on a 4 m- or 1 m-class telescope. (MUSE is discussed later in this chapter;
Bacon et al., 2004).
In addition to the basic ingredients listed above, the most common merit functions are
the grasp, the specific grasp, A × dΩ (how much is grasped within each spatial resolution
element of the instrument), and etendue, A × Ω × ϵ, where ϵ is the total system efficiency
from the top of the atmosphere to the detected photoelectron. Etendue is more funda-
mental than grasp since high-efficiency instruments are the true performance engines.
3D spectroscopy instrumentation 89
Despite the fact that an instrument with an unreported efficiency is much like a car sans
fuel-gauge or speedometer, recovering ϵ from the literature is often not possible. For this
reason we resort to grasp, but note that in some cases this gives an unfair comparison
between instruments.
If there is no premium on spatial information then ‘spectral power’, R × NR , is suit-
able. At the opposite extreme, where spatial information is paramount, a suitable merit
function is A × dΩn × NΩ = A × dΩn −1 × Ω, where n = 1 for high specific grasp and −1
for high resolution. In the context of 3D spectroscopy, merit functions that combine
spatial and spectral power are appropriate: Ω × R, A × Ω × R × ϵ, or their counterparts
replacing Ω with dΩ. If any information will do, NR × NΩ alone gives a good synopsis
of the instrument power since this effectively gives the number of resolution elements
(related to detector elements) that have been effectively utilized by the instrument.
An attempt at a grand merit function can be formulated by asking the following,
sweeping question: ‘How many resolution elements can be coupled efficiently to the largest
telescope aperture (A) covering the largest patrol field (Ωs ) for as little cost as possible?’
In this case, the figure of merit may be written:
F OM = ϵ × (∆λ/λ) × (Ω/dΩ) × A × Ωs × £−1 (3.1)

= ϵ × NR × NΩ × A × Ωs × £ −1
(3.2)
where ∆λ is the sampled spectral range, and £ is the cost in the suitable local currency.
To this figure of merit one may add the product Rn × dΩm , where n, m = 1 if resolution is
science-critical in the spectral and spatial domains (respectively), n, m = −1 if coverage
is science-critical, or n, m = 0 if resolution and coverage are science-neutral (in which
case you are not trying hard enough!).
From this discussion it is clear that a suitable choice of merit function is complicated,
and must be science driven. The relative evaluation of instruments cannot be done sensi-
bly in the absence of a science-formulated FOM; the outcome of any sensible evaluation
will therefore depend on the science formulation. For this reason, when we compare
instruments we strategically retreat and explore the multi-dimensional space of the fun-
damental parameters of spatial resolution, spectral resolution, specific grasp, total grasp,
spectral power, and NR versus NΩ .
3.1.3 Why spectral resolution is so important

In addition to the intrinsic merits and requirement of high spectral resolution for certain
science programmes, high resolution is of general importance for improving signal-to-
noise (S/N) in the red and near-infrared (NIR). For ground-based observations, terres-
trial backgrounds from 0.7–2.2 microns suffer a common malady of being dominated by
extremely narrow (m s−1 ) airglow lines, typically from OH molecules. Unlike the thermal
IR, however, there is a cure to lower the background without going to high altitude or
space. The airglow lines cluster in bands, and the lines within the bands may be separated
at R = 3000–5000. This means that at these resolutions, while the mean background level
within the spectral bandpass is constant, the median drops precipitously: more spectral
resolution elements are at lower background level in interline regions. The lines them-
selves, however, remain unresolved until R ∼ a few × 105 , so that above R = 4000 one
continues to increase the fraction of the spectral bandpass at low-background levels.
As an illustration, we show the terrestrial sky background in a spectral region at
0.8 microns observed by D. York and J. Lauroesch (private communication) with the
KPNO (Kitt Peak National Observatory) 4 m echelle. In Figure 3.2 the sky spectra,
observed at an instrumental resolution of 33,000, is degraded to illustrate the resulting
Figure 3.2. Night sky near 0.8 µm at 250 < R < 33,000.
Figure 3.3. Cumulative distribution of resolution elements as a function of the background

level proportional to S/N (increasing to the left) for 250 < R < 33,000 (labelled).
change in the distribution of background levels. In Figure 3.3, the normalized, cumu-
lative distribution of resolution elements as a function of background level is plotted
for different instrumental resolutions. For background-limited measurements, the S/N is
proportional to the inverse square root of the background level. Hence the median back-
ground level gives an effective scaling for sensitivity gains with spectral resolution. It
can be seen that the largest changes in the median background level occur between
1000 < R < 4000, but significant gains continue at higher resolution. The result can
be qualitatively generalized to other wavelengths in the 0.7–2.2 µm regime. While the
lines become more intense moving to longer wavelengths, the power spectrum (in wave-
length) of the lines appears roughly independent of wavelength in this regime (compare
Figure 3.4. Maximum spectral resolution versus telescope diameter to stay background (versus
detector) limited for different assumptions of instrument efficiency (ϵ) and spatial sampling (dΩ).
The solid line assumes ϵ = 0.15 and dΩ = 1 arcsec2 .
Maihara et al., 1993; Hanuschik, 2003). Note this is a qualitative assessment that should
be formally quantified.
3.1.4 The detector limit II: read noise

Our infatuation with spectral resolution is a problem given the modern predilection for
high angular resolution. After the Hubble Space Telescope there is no turning back! There
is, however, a limit, due to detector noise, which we always want to be above. The goal
is to be photon-limited (either source or background) because this is fundamental (it is
the best we can do), and for practical purposes, S/N is independent of sub-exposure time
and detector sampling.
To stay photon-limited in the background-limited regime puts significant constraints
on the Ω–R sampling unit. The spatial and spectral sampling unit cannot be too fine for
a given A and ϵ. For 8 m- and 4 m-class telescopes we calculate:
R/dΩ < 16, 500(DT /9m)2 (t/1h)(ϵ/0.15) arcsec−2 , or (3.3)
2
< 2500(DT /3.5m) (t/1h)(ϵ/0.15) arcsec −2
(3.4)
where DT is the telescope aperture diameter and t the (single destructive-read) exposure
length. The general case is shown in Figure 3.4. To reach spectral resolutions well above
R = 5000, which is advantageous for background reduction, a telescope significantly in
excess of 10 m is needed for apertures significantly under 1 arcsec−2 .
With these considerations in mind, in the next three sections (Section 3.2–3.4) we turn
to approaches and examples of existing instruments, followed by three sections (Sections
3.5–3.7) in which we summarize the range of these instruments, what parameter space
is undersampled, and the prospects for future instruments. Throughout, we attempt
to provide relatively complete instrument lists. No doubt, some instruments have been
camera focal length = f2

w’
grating−collimator distance = d_gc

D2
G.N.
β
α
w D1
grating size = Wg
collimator focal length = f1
collimator−grating distance = d_cg
Figure 3.5. Basic spectrograph layout schematic for reflective/refractive collimator, reflection
grating and refractive camera.
overlooked, and also the field of instrumentation advances rapidly. Reports of additional
instruments or corrections are welcome.
3.2 Grating-dispersed spectrographs

Basic spectrograph theory and design can be found in most standard optics textbooks. Of
particular note is the excellent monograph on astronomical optics by Schroeder (2000).
In Section 3.2.1 we summarize the salient features to provide a consistent nomenclature,
and to put these features into context of our discussion of 3D spectroscopy, specifically
what drives consideration of merit functions that tune spatial versus spectral perfor-
mance. The balance of this section includes a description of dispersive elements (Section
3.2.2), coupling methods and modes (Section 3.2.3–3.2.11), and summary considerations,
including a discussion of sky-subtraction problems and solutions (Section 3.2.12).
3.2.1 Basic spectrograph design

In a 3D spectrographic system, there is a premium on packing spatial information
on to the detector. To achieve sufficient spectral resolution at the same time requires
balancing the trades between system magnification and dispersion. Starting with the
grating equation, generalized for a grating immersed in medium of index n: mλ =
nΛg (sin β + sin α), where Λg is the projected groove separation in the plane of the
grating, m the order, and α and β the incident and diffracted grating angles relative
to the grating normal in the medium, we can write the angular and linear dispersion
as γ ≡ dβ/dλ = m/nΛg cos β = (sin β + sin α)/λ cos β, and dl/dλ = f2 γ. Figure 3.5
illustrates a basic spectrograph, defines these angles and subsequent terms.
The system magnification can be broken down into spatial and anamorphic factors.
The physical entrance aperture width, w, is re-imaged on to the detector to a physical
width w′ , demagnified by the ratio of camera to collimator focal lengths. Hence the
spatial width (perpendicular to dispersion) is given as wθ′ = w(f2 /f1 ). For non-imaging
feeds (i.e. fibres or lenslets), it is advantageous to pack as much information as possible
into a given pixel, as long as individual spatial entrance elements can be resolved. This
means cameras must be as fast as possible relative to their collimators. For imaging feeds
(slits or slicers), the desire to preserve and sample the spatial information retained in the
slit means the choice must be science driven.
In the dispersion direction, an additional, anamorphic factor, r, arises due to the fact
that grating diffraction implies incident and diffracted angles need not be the same. Hence
incident and diffracted beam sizes scale as r = D1 /D2 = cos α/ cos β. This arises because
in general A × Ω is conserved; if the beam gets larger, the angles get smaller. Another
way to think of this is in terms of the definition of r = |dβ/dα|, and ask for a given dα
(angular slit width) what is dβ such that dλ = 0? This result can then be derived from
the grating equation. In any case, β/α < 1 implies magnification, while β/α > 1 implies
demagnification. The re-imaged slit width in the spectral dimension is then wλ′ = rwθ′ . In
Littrow configurations, important below, α = β = δ (the latter being the grating blaze
angle), and so there is no anamorphic factor. Since the re-imaged slit width always
degrades the instrumental spectral resolution, it is always advantageous, in this sense,
to have anamorphic demagnification. However, depending on the pixel sampling, optical
aberrations, and slit size, wλ′ may not be the limiting factor in instrumental resolution.
Anamorphic demagnification also comes at a cost: the camera must be large (larger than
the collimator) to capture all of the light in the expanded beam. Demagnification never
hurts resolution, but the cost should be weighed against the gains.
The spectral resolution can now be written as R = λ/dλ, or R = λ(γ/r)(f1 /w). The
term γ/r indicates that we want large dispersion, but that we can get resolution also from
anamorphic demagnification. The terms f1 /w indicate that we want a long collimator at
fixed camera focal length, requiring a field lens or white-pupil design to avoid vignetting.1
Alternatively, we may rewrite the equation as R = λ(γ/r)(D1 /θDT ), noting θ is the angle
on the sky, dλ = wλ′ /(dl/dλ), w = fT θ, and f1 /d1 = fT /DT , where fT /DT refer to the
effective focal ratio of whatever optics feed the spectrograph, e.g. the telescope. The
combination of r and D1 indicates that we want a larger collimator and an even larger
camera. Using the grating equation we may write R = (f1 /w)(sin β + sin α)/ cos α, which,
in Littrow configurations, reduces to R = (f1 /w)2 tan α. In the latter situation it is clear
that resolution can be dispersion-driven by going to large diffraction angles, α, which
requires large gratings.
3.2.2 Dispersive elements

We distinguish here principally between reflection and transmission gratings. Trans-
mission gratings yield much more compact spectrograph geometries. This leads to less
vignetting and better performance with smaller optics.
1
A field lens, which sits near a focus to avoid introducing power into the beam, serves to
move the spatial pupil to a desirable location in the system. This is often the grating, but in
general can be the location such that the overall system vignetting is minimized. A white-pupil
design (Baranne, 1972; Tull et al., 1995) is one that re-images a pupil placed on a grating,
typically on to a second grating (e.g. a cross-disperser) or the camera objective. It is ‘white’
because the pupil image location is independent of wavelength even though the light is dispersed.
Reflection gratings
Reflection gratings come in three primary varieties: ruled surface-relief (SR), holograph-
ically etched SR, or volume-phase holographic (VPH). We list the pros and cons of each
of these.
Ruled SR gratings have the advantage of control over the groove shape, blaze and
density, which provides good efficiency in higher orders (e.g. echelle) at high dispersion.
There are existing samples of masters with replicas giving up to 70% efficiency, but 50–
60% efficiency is typical, with 40% as coatings degrade. Scattered light and ruling errors
can be significant, and existing masters are limited in type and size. It does not appear
to be possible to make larger masters with high quality at any reasonable cost.
Holographically etched SR gratings have low scattered light, the capability to achieve
high line density (hence high dispersion) and large size. However, they have low efficiency
(< 50%) because symmetric grooves put equal power in positive and negative orders.
VPH gratings can be made to diffract in reflection (Barden et al., 2000), but have
not yet been well developed for astronomical use. Reflection gratings can be coupled to
prisms to significantly enhance resolution via anamorphing (Wynne, 1991).
Transmission gratings
Transmission gratings are either SR or VPH and, when coupled with prisms, are referred
to as grisms. SR transmission gratings and grisms are efficient at small angles and low
line densities (good for low-resolution spectroscopy), but are inefficient at large angles
and line densities due to groove shadowing. Transmission echelles do exist, but have 30%
diffraction efficiencies or less.
VPH gratings and grisms are virtually a panacea. They are efficient over a broad range
of line densities and angles. Any individual grating is also efficient over a broad range of
angles (what is known as a broad ‘superblaze’; see below). Peak efficiencies are as high as
90%; they are relatively inexpensive to make and likewise to customize; and they can be
made to be very large (as large as your substrate and recording beam, now approaching
0.5 m). Their only disadvantages are that they have, to date, been designed for Littrow
configurations.
It is worth dwelling somewhat on the theory and subsequent potential of VPH grat-
ings. There still remain manufacturing issues of obtaining good uniformity over large
areas (Tamura et al., 2005), but it is reasonable to be optimistic that refinement of the
process will continue at a rapid pace. Application in the NIR for cryogenic systems is
also promising: coefficient thermal expansion mismatch between substrate and diffracting
gelatin, potentially causing delamination, does not appear to be a concern (W. Brown,
private communication, this Winter School). Blais-Ouellette et al. (2004) have confirmed
that diffraction efficiency holds up remarkably well at 77 K, but that the effective line den-
sity changes with thermal contraction. We can expect most grating-fed spectrographs in
the future will use VPH gratings alone or in combination with conventional (e.g. echelle)
gratings. The capabilities of VPH gratings will open up new design opportunities, many
of which will be well suited to 3D spectroscopy.
3.2.3 VPH grating operation and design

Diffraction arises from modulation of the index of refraction in a sealed layer of thickness d
of dichromated gelatin (the material is hygroscopic), with mean optical index n2 . Typical
values for n2 are around 1.43, but the specific value depends sensitively on the modulation
frequency (i.e. the line density Λ) and amplitude, ∆n2 , and the specifics of the exposure
and developing process. (Note that it is not currently possible to predict the precise
value of n2 from a manufacturing standpoint.) The seal is formed typically by two flat
substrates, but this can be generalized to non-flat surfaces and wedges (i.e. prisms).
Because this layer represents a volume (d ≫ λ), the diffraction efficiency is modulated
by the Bragg condition: α = β. These angles are defined here with respect to the plane
of the index modulations.
The wonder of VPH gratings is the ability to custom-design them. Starting with a
science-driven choice of dispersion and wavelength, the grating equation and dispersion
relation given the Bragg condition uniquely set the line frequency and angle, respectively,
for unblazed gratings. The key to high diffraction efficiency is then to tune the gelatin
thickness and index modulation amplitude such that diffraction efficiency is high in both
s- and p-polarizations (the s-polarization electric vector is perpendicular to the fringes).
This can be done by brute force via rigorous coupled wave calculations, or by noting that
in the so-called ‘Kogelnik limit’ the diffraction efficiencies are periodic in these quantities
(Barden et al., 2000; Baldry et al., 2004). The two polarizations have different periodici-
ties, i.e. VPH gratings are in general highly polarizing, so the trick is finding the (d, ∆n2 )-
combination that phases one pair of s and p efficiency peaks. Thinner gel layers yield
broader bandwidth over which the diffraction efficiency is high, relative to the efficiency
at the Bragg condition. The thinner the layer, the larger the index modulation required to
keep the efficiency high in an absolute sense. Modulations above 0.1 are very difficult to
achieve, and more typical values are in the range of 0.04–0.07; gel layers are in the range
of a few to a few tens of microns. In practice, because there is limited manufacturing
control over the index modulation and effective depths of the gelatin exposure, gratings
requiring very precise values in these parameters will be difficult to make and have large
inhomogeneities. Our experience is that it is useful to understand how wavelength and
resolution requirements can be relaxed to locate more robust design parameters.
Blazed VPH gratings

Nominally, the fringe plane is parallel to the substrate normal (indicated by the angle
φ = 0). This yields an unblazed transmission grating. Essentially all astronomical VPH
gratings in use are made this way. There is concern that tilted fringes will curve with
the shrinkage of the gelatin during development (Rallison and Schicker, 1992), but this
concern has not been fully explored. By tilting the fringes (this is done simply by tilting
the substrate during exposure in the hologram), one can enter several different interesting
regimes, as illustrated concisely by Barden et al. (2000) in their Figure 1: small |φ| yields
blazed reflection gratings, φ = 90 deg produces unblazed reflection gratings, and large
|φ| blazes the reflection gratings. ‘Large’ and ‘small’ depend on the angle of incidence,
as illustrated below. The sign convention is such that positive φ decreases the effective
incidence angle. The incident and reflected angles in the gelatin, α2 and β2 , are related
by α2 = β2 + 2φ, with α2 − φ being the effective diffraction angle. The grating equation,
when combined with the Bragg condition yields: mλb = 2n2 Λ sin(α2 − φ), where λb is the
Bragg wavelength, and Λ = Λg cos φ is the fringe spacing perpendicular to the fringes.
We use Baldry et al.’s (2004) nomenclature; their Figure 1 is an instructive reference for
this discussion.
Baldry et al. (2004) work out the case for no fringe tilt with flat or wedged substrates.
Here we give the case of flat substrates but arbitrary φ. Burgh et al. (2007) extend this to
include arbitrary fringe tilt (φ). The relevant angles with respect to the grating normal
can be found with these equations in terms of the physical grating properties:
% ) * &
sin α
sin α = n2 sin α2 , and sin β = n2 sin sin−1 − 2φ (3.5)
n2
R
A
Figure 3.6. Resolution merit function and anamorphic factor for blazed VPH gratings with
mean gel index n2 = 1.43. Typical SR gratings have 1.05 < 1/r < 1.2.
The anamorphic factor and dispersion are still defined in terms of α and β as given in
Section 3.2.1. With the interrelation of these angles as given above, it is easy to show
the logarithmic angular dispersion at the Bragg wavelength is:
% ) * &-
sin α
dβ/dlogλ = 2n2 cos φ sin sin −1
−φ cos β. (3.6)
n2
To understand the potential advantages of blazed transmission gratings, we define

a resolution merit function as 1r dβ/dlogλ, i.e. the product of the logarithmic angular
dispersion and the anamorphic factor. With this function we can explore, in relative
terms, whether tilting the fringes yields resolution gains. Figure 3.6 shows the anamorphic
factor and the resolution merit function versus grating incidence angle for positive and
negative fringe tilts. Negative fringe tilts give a small amount of increased resolution
at a given α by significantly increasing dispersion, which overcomes an increase in the
anamorphic magnification (mag). This means the detector is less efficiently used. Negative
fringe tilts also limit the usable range of α for which β < 90 deg (transmission), and hence
the maximum achievable resolution in transmission that can be achieved is lowered with
negative fringe tilts.
With positive fringe tilts, the anamorphic demagnification increases strongly at large
incidence angles, although there is little gain in going to φ > 15 deg. Note that the demag-
nification becomes < 1 (i.e. magnification) roughly when α ∼ 1.5 φ. This is when the
effective diffraction angle (α2 − φ), changes sign with respect to the tilted fringes (the
grating remains in transmission). The overall resolution decreases with increased positive
fringe tilt, but the decrease is modest for small tilt angles. Given the large increase in
anamorphic demagnification relative to the modest loss in resolution, for small tilt angles
there is a definite gain in information: A +5 deg tilt gives a 12% loss in the resolution
a) Unblazed, broad-band grating d) Unblazed, narrow-band grating
b) Double grating: high-dispersion mode e) Blazed, broad-band grating
c) Double grating: NB filter mode f) Triple-notch grating
Figure 3.7. Novel grating modes. a) Conventional broad-band application now becoming a
staple of modern spectrographs. b) Double-grating geometry yielding a net dispersion of ∼ the
sum of the two individual grating dispersions (gratings are not necessarily identical, but angles
must be adjusted accordingly). c) Double-grating geometry yielding a narrow-band filter with
field-dependent bandpass given by the Bragg condition (gratings are identical) d)–f) Narrow
bandpass gratings unblazed (d), blazed (e), and combined (f) to form a notch grating. Other
modes are discussed in the text. Panels a), d)–f) show both the grating configuration as well as
a cartoon-sketch of the diffraction efficiency as a function of location of the detector dispersion
axis, labelled for the mean wavelength regime of the diffraction bandpass.
merit function at α = 60, but a 51% gain in the anamorphic demagnification. With suit-
ably good optics and detector sampling the demagnified image, this equates directly into
an increase in the number of independent spectral resolution elements, replete with a 72%
increase in spectral coverage. The loss in resolution can easily be made up by slightly
increasing α (in this case, from 60 to 63 deg) and modulating Λ in the grating design
to tune the wavelength. Instruments with blazed, high-angle VPH gratings with tilts of
5 < φ < 15 deg will allow for the high resolution needed to work between sky lines, while
efficiently packing spectral elements on to the detector. This is critical in the context of
3D spectroscopy, where room must also be made for copious spatial elements.
Unusual VPH grating modes

In addition to tilted fringes, VPH gratings pose opportunities for a number of novel
modes well suited to 3D spectroscopy. Figure 3.7 illustrates some of these. With very
high diffraction efficiency it is now reasonable to consider combining gratings to augment
the dispersion, and hence resolution. If the two gratings are kept parallel but offset along
the diffraction angle, they can serve as (tunable) narrow-band filters, an alternative
to etalons (e.g. Blais-Ouellette et al., 2006). Barden et al. (2000) have explored using
multiple gelatin layers with different line frequencies to select Hα and Hβ in separate
bandpasses. By slightly rotating one set of lines, sufficient cross-dispersion is added to
space the two spectra – one above the other – on the detector. This is well suited for
spectrographs fed with widely spaced fibres or slitlets (i.e. an underfilled, conventional
long-slit spectrograph), and represents an interesting trade-off in wavelength and spatial
multiplex. At sufficiently high dispersion (and hence limited bandpass), the number of
layers could be increased to mimic a multi-order echelle.2 The advantage of this approach
is in resolution and wavelength coverage.
An alternative approach is something we refer to as ‘notch’ gratings. Here, we take
advantage of the relative ease (from a manufacturing standpoint) of achieving a narrow
bandpass and combine gel layers tuned to different, non-overlapping wavelength band-
passes at a given incidence angle (e.g. by changing the line frequency). By also tuning the
fringes with modest tilts, each bandpass can be centred on a different, non-overlapping
portion of the detector. Bandpasses will have to be carefully crafted by tuning grating
parameters to avoid parasitic contamination in the other bands. The figure illustrates
positive and negative tilts, but the tilts could be arranged all to be positive to take
advantage of the anamorphic factors described above. This offers another way to slice
the datacube, one which allows for sparse spectral sampling of key spectral diagnostics
over a broad wavelength range (e.g. [OII]λ3727, Hβ+[OIII]λλ4959,5007 and Hα) at high
dispersions, with ample room left over on the detector for significant spatial multiplex.
3.2.4 Summary of implications for 3D spectrograph design

The most compact spectrograph designs yield the highest-efficiency, wide-field systems
needed to grapple with attaining large angular coverage for 3D spectroscopy. To also
obtain high enough spectral resolution to work between the atmospheric airglow often
requires significant dispersive power and anamorphic demagnification. Large anamor-
phic demagnification, while not free (larger camera optics are required), is well-suited
to packing information on to the detector. This is particularly important in 3D appli-
cations where spatial multiplex is at a premium. VPH transmission gratings are clearly
preferred because they lend themselves to compact spectrograph geometry and provide
high diffraction efficiency. We have shown they can, in principle, also yield large anamor-
phic demagnification. With high-angle, double, and blazed VPH gratings, echelle-like
resolutions can be achieved at unprecedented efficiency (75–90% in diffraction alone).
Unusual modes to produce tunable narrow-band filters and notch gratings also open up
the possibility for well-targeted sparse, spectral sampling.
3.2.5 Coupling formats and methods: overview

The essence of the 3D spectrometer lies in the coupling of the telescope focal plane to
the spectrograph. We review the four principal methods: (i) direct fibres, (ii) fibres +
lenslets, (iii) image-slicers, and (iv) lenslet arrays, or pupil-imaging spectroscopy. A nice,
well-illustrated overview can be found in Allington-Smith and Content (1998); additional
discussion of the merits and demerits of different approaches can be found in Alighieri
et al. (2005). Here we also make an evaluation. We discuss a fifth mode not seen in the
literature, which we refer to as (v) ‘filtered multi slits’. Many spectrographs either have,
or could easily be modified to have, this capability. We also describe (vi) multi-object
configurations: a mode which will undoubtedly become more common in the future.
Throughout this discussion, we distinguish between near-field versus far-field effects.
The near-field refers to the light distribution at the focal surface, e.g. fibre ends, and what
is re-imaged ultimately on to the CCD. The far-field refers to the ray-bundle distribution,
i.e. the cross-section intensity profile of the spectrograph beam significantly away from
the focal surface. Different coupling methods offer the ability to remap near- and far-field
light-bundle distributions, which can have advantages and disadvantages.
2
A true cross-dispersed echelle-like grating would work, in principle, with two layers, rotated
by 90 degrees. VPH gratings have not yet been made with high efficiency in multiple orders, but
see Barden et al. (2000) for measurements up to order 5.
3.2.6 Direct fibre coupling
The simplest and oldest of methods consists of a glued bundle of bare fibres mapping
the telescope to spectrograph focal surfaces. With properly doped, AR-coated fibres
throughput can be at or above 95%, which can be compared to 92% reflectivity off one
freshly coated aluminium surface. These have the distinct advantage of low cost and high
throughput. As with all fibre-based coupling, there is a high degree of flexibility in terms
of reformatting the telescope to spectrograph focal surfaces (for example, it is easy to mix
sky and object fibres along slit), and the feeds can be integrated into existing long-slit,
multi-object spectrographs. However, bare fibre integral field units (IFUs) are not truly
integral and do not achieve higher than 60–65% fill factors (on the deleterious effects
of buffer stripping of small fibres, see Oliveria et al., 2005). This coupling is perhaps
the most cost-effective mode for cases where near-integral sampling is satisfactory, and
preservation of spatial information is not at a premium.
Information loss and stability gain with fibres

Focal ratio degradation (FRD) and azimuthal scrambling represent information loss (an
entropy increase). FRD specifically results in a faster output f -ratio (Ramsey, 1988).
This has an impact on spectrograph design or performance since either the system will
be lossy (output cone overfills optics), or the spectrograph has to be designed for the
proper feed f -ratio. PMAS (Potsdam Multiaperture Spectrophotometer; Roth et al.,
2005) is an excellent example of how to properly design a spectrograph to handle fast
fibre-output beams. The existing WIYN (Wisconsin–Indiana–Yale) Bench spectrograph
is a good example of how not to do it. In fact, it is so bad that we rebuilt it (Bershady
et al., 2008); we were able to recapture 60% of the light (over a factor of 2 in throughput)
with no loss of spectral resolution in the highest resolution modes.
Azimuthal scrambling can help and hurt. While scrambling destroys image information,
it symmetrizes the output beam, ameliorating, to some extent, the effect of a changing
telescope pupil on HET-like (Hobby–Eberly Telescope) or SALT-like (Southern African
Large Telescope) telescopes by homogenizing the ray bundle. Thus, the contribution of
spectrograph optical aberrations to the final spectral image is more stable. (This is a
far-field effect.)
Radial and azimuthal scrambling together homogenize near-field illumination, e.g. the
seeing-dependent slit function is decreased. Radial scrambling and FRD are one and the
same (compare Ramsey, 1988 and Barden et al., 1993), so that one trades information
loss for stability (similar to the trade of precision for accuracy). In practice, fibre-input
beam speeds of f /3 (PMAS) to f /4.5 (HET and SALT) are desirable. However, with
fast input/output f -ratios this limits possible spectrograph demagnification since it is
expensive to build faster than f /2 for large cameras.
Telecentricity
Because azimuthal scrambling symmetrizes a beam, if the input light-cone is misaligned
with the fibre axis, the output beam (f -ratio) is faster. This is not FRD. To avoid this
effect, fibre telecentric alignments of under a degree are needed even for f -ratios as fast
as 4–6 (Bershady et al., 2004; Wynne and Worswick, 1989).
Causes of FRD
Excessive FRD in fibres is due to stress. Hectospec (Fabricant et al., 2005) embodies
an excellent example of how to properly treat fibres and fibre cabling (Fabricant et al.,
1998; see also Avila et al., 2003 in the context of FLAMES on VLT). Fibre termination
Figure 3.8. Output fibre irradiance (encircled energy versus beam f -ratio) for fibre cables on
the WIYN Bench Spectrograph. The input beam profile is an unapodized f /6.3 beam with an
f /17 central obscuration (labelled). Output beam profiles are faster, due to FRD, and are well
fitted by a Sersic model of index 1/n = 5 (S. Crawford, private communication).
and polishing can also induce stress. Bershady et al. (2004) discuss some other IFU-
related issues in terms of buffering. However, even for perfectly handled fibres, there is
internal scattering, the cause of which has long been a debate. Nelson (1988) suggested
a combination of (i) Rayleigh scattering (variation in fibre refractive index); (ii) Mie
scattering (fibre inhomogeneities comparable to the wavelength); (iii) stimulated Raman
and Brillouin scattering (not relevant at low signal level in astronomical applications);
and (iv) micro-bending. Micro-bending seems a likely culprit; it is the unsubstantiated
favourite in the literature. Micro-bending models predict a wavelength-dependent FRD.
While Carrasco and Parry (1994) tentatively see such an effect, neither Schmoll et al.
(2003) nor Bershady et al. (2004) confirm the result. However, these studies use different
measurements methods. More work is required to understand the physical cause(s) of
FRD, and with this understanding, perhaps, reduce the amplitude of the effect. We find
FRD produces an output fibre beam profile which can be well-modelled by a Sersic
function (Figure 3.8; S. Crawford, private communication). This says something about
either the scattering model or how seriously to take physical interpretations of Sersic-law
profiles of galaxies.
Quality versus quantity

Fibres offer the opportunity of easily trading quality for quantity in terms of packing the
spectrograph slit. Scattered light within the spectrograph, combined with fibre azimuthal
scrambling means that spatial information in the telescope focal plane is coupled to all
adjacent fibres in the slit. Closely packing fibres in the slit can make clean spectral extrac-
tion difficult. The WIYN Bench Spectrograph is a good example where the amplitude of
scattered light is low, fibre separation is large and ghosting is negligible. This spectro-
graph and feeds are optimized for clean extraction with little cross-talk (1% cross-talk in
the visible in optimum S/N aperture, degrading to 10% in the NIR). Information packing
in the spatial dimension is modest due to fibre separation, while information packing
Figure 3.9. Critical sampling with densely packed fibres.
in the spectral dimension is high due to large anamorphic factors. Other systems have
significant spectral overlap. For example, staggered slits, where fibres are separated by
only their active diameter (COHSI; Kenworthy et al., 1998) make it difficult to extract a
clean spectrum and optimize S/N at the same time, but the spatial multiplex is increased.
There is no one right answer, but definitely a decision worthy of science-based consid-
eration.
Image reconstruction and registration

Even without lenslets, densely sampled fibres provide excellent image reconstruction on
spatial scales comparable to the fibre diameter. One can achieve the theoretical sampling
limit with a three-position pattern of half-fibre-diameter dithers (Figure 3.9; compare
Koo et al., 1994 in the context of under-sampled HST/WFPC-2 data). Even with sparse
sampling, registration of the spectral datacube with broad-band images can be achieved
to 10% of the fibre diameter by cross-correlating the spectral continuum with respect to
broad-band images or integrated radial light profiles (Bershady et al., 2005). Kelz et al.
(2006) show how it is possible to reproduce the continuum image of UGC 463 using the
PPak fibre bundle without any subsampling.
Summary of instruments
Some of the first IFUs were on the KPNO 4 m RC spectrograph: DensePak-1 followed by
DensePak-2 (Barden and Wade, 1988; for other early IFUs, see also Guerin and Felenbok,
1988). The last incarnation (Barden et al., 1998) was on WIYN. Conceptually, these
instruments spawned SparsePak (WIYN; Bershady et al., 2004) and PPak (PMAS, Calar
Alto; Verheijen et al., 2004; Kelz et al., 2006). A more versatile single instrument suite,
built for the William Herschel Telescope (WHT), is INTEGRAL (WYFOS, Wide-Field
Optic Spectrograph), which offers several plate-scales and formats (Arribas et al., 1998),
and a sophisticated and well-thought-out mapping between telescope and spectrograph
focal planes. These are all shown in Figure 3.10. GOHSS is one case of a NIR (0.9–1.8 µm)
application (Lorenzetti et al., 2003). VIRUS (Hill et al., 2004) and APOGEE (Allende
Prieto et al., 2008) are the only planned future instruments. A summary of existing and
future optical and NIR direct fibre-fed IFU spectrographs are listed in Table 3.1.
3.2.7 Fibre + lenslet coupling

The basic concept of lenslet coupling to fibres is again, as with bare fibres, to remap a
2D area in the telescope focal surface to a 1D slit at the spectrograph input focal surface.
The key difference is in the fore-optics, which consists of a focal expander and lenslet
array; these feed the fibre bundle. The focal expander matches to the scale of the lenslet
array. Allington-Smith and Content (1998) and Ren and Allington-Smith (2002) present
some technical discussion and illustration of the method. Each micro-lens in the array
then forms a pupil image on the fibre input face. The pupil image is suitably smaller
than the lenslet to allow the fibres to be packed behind the integral lenslet array. This
reduction speeds up the input beam (A × Ω is conserved). Given the previous discussion
concerning FRD, this can be advantageous to minimize entropy increase.
Table 3.1. Direct fibre-coupled integral field instruments
DT Ω dΩ
Instrument Tel. (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
Existing optical instruments
DensePak WIYN 3.5 564 6.2 91 1.02 1000 1024 0.04

3.5 564 6.2 91 0.07 13750 1024 0.04
3.5 564 6.2 91 0.04 24000 1024 0.04
3.5 119 1.3 91 1.02 1000 1024 0.04
3.5 119 1.3 91 0.07 13500 1024 0.04
3.5 119 1.3 91 0.04 24000 1024 0.04
SparsePak WIYN 3.5 1417 17.3 82 1.02 800 819 0.07
3.5 1417 17.3 82 0.07 11000 819 0.07
3.5 1417 17.3 82 0.03 24000 819 0.07
PPak CA 3.5 2070 5.64 367 0.15 7800 1183 0.15
INTEGRAL WHT 4.2 32.6 0.159 205 0.22 2350 515 ···
4.2 32.6 0.159 205 0.94 550 515 ···
4.2 139.3 0.64 219 0.22 2350 515 ···
4.2 139.3 0.64 219 0.94 550 515 ···
4.2 773 5.73 135 0.07 2350 300 ···
4.2 773 5.73 135 0.90 550 300 ···
Future optical instruments
VIRUS HET 9.2 32604 1.0 32604 0.505 811. 410 0.16
Existing near-infrared instruments
GOHSS TNG 3.6 44.2 1.77 25 0.12 4380. 512 0.13
Future near-infrared instruments
APOGEE SDSS 2.1 942 3.14 300 0.10 23500 2350 0.22
Figure 3.10. Direct-fibre IFUs on optical spectrographs. The top row shows the legacy started
by S. Barden with DensePak-1 and DensePak, leading to SparsePak, PPaK on the KPNO 4 m,
WIYN, and Calar Alto, respectively. The bottom row shows Hexaflex and INTEGRAL on WHT
with their multiple, selectable bundles and ample sky-fibres.
At the output stage, the option exists to reform the (now azimuthally scrambled) slit
image with an output micro-lens linear array, or to use bare fibres. Without lenslets,
the input f -ratio to the spectrograph will be faster, which means there is less possibility
for geometric demagnification via a substantially faster camera. In this case the spec-
trograph also re-images the fibre-scrambled telescope pupil: the image varies with tele-
scope illumination, while the ray bundle distribution (far-field) varies with the telescope
image.
The positive attributes of lenslet-fed fibre arrays are: (i) improved filling factors to
near unity; and (ii) control of input and output fibre f -ratio. The latter permits effective
coupling of a slow telescope f -ratio to fibre input at a fast, non-lossy beam speed and like-
wise permits effective coupling of fibre output to spectrograph. The negative attributes of
this coupling method include (iii) increased scattered light (from the lenslet array); (iv)
lower throughput (due to surface reflection, scattering, and misalignment). For example,
typical lenslet + fibre units yield only 60–70% throughput (Allington-Smith et al., 2002).
When there is a science premium on truly integral field sampling, the above two factors
do not outweigh the filling factor improvements. Finally, there is the more subtle effect of
whether or not to use output lenslets. Aside from the matter of f -ratio coupling, there is
the issue of whether swapping the near- and far-field patterns is desirable for controlling
systematics in the spectral image. It amounts to assessing whether the spectrograph is
‘seeing-limited’, i.e. limited by spatial changes in the light distribution within the slit
image formed by the fibre and lenslet, or aberration limited.
Prime examples of optical instruments on 8 m-class telescopes include VIMOS
(Le Fevre et al., 2003), GMOS (Gemini-N,S, Allington-Smith et al., 2002), and
FLAMES/GIRAFFE in ARGUS or multi-object IFU modes (Avila et al., 2003).3 Typ-
ical characteristics of these devices are fine spatial sampling (well under an arcsec) and
modest spectral resolution. ARGUS is an exception, achieving resolutions as high as
39,000. Its multi-object mode is also unique, and powerful (see later discussion). On 4–
6 m-class telescopes there are PMAS (Roth et al., 2005), Spiral+AAOmega (Kenworthy
et al., 2001; Saunders et al., 2004), MPFS (Afanasiev et al., 1990)4 and IMACS-IFU
(Schmoll et al., 2004).5 Compared to most direct-fibre IFUs on comparable telescopes,
these instruments also have finer spatial sampling.
NIR instruments include SMIRFS (Haynes et al., 1999), and COHSI, which is a pre-
cursor – in some regards – to CIRPASS (Parry et al., 2004). An interesting application
of flared fibres is discussed by Thatte et al. (2000) for cryogenic systems.
A summary of existing and future optical and NIR lenslet + fibre coupled IFU spec-
trographs is listed in Table 3.2. While it may seem surprising that no future instruments
appear to be planned, we will discuss one possible instrument for the 30 m Telescope
(TMT) below.
3.2.8 Slicer coupling

Image-slicers have been around for a long time, primarily serving the high-resolution
community, e.g. to slice a large fibre into a thin, relatively short slit to feed cross-dispersed
echelles (for one recent example, see Tull et al., 1995) Extending the concept into a
3D mode follows the same basic notion, which can be thought of as deflecting slices
of the telescope image plane both along and perpendicular to the slice through a pair
of reflections. These reflections have power to reform the focal-plane image. Given the
3
See also www.eso.org/instruments/flames/inst/Giraffe.html.
4
See also www.sao.ru/hq/lsfvo/devices/mpfs/mpfs main.html.
5
See also www.ociw.edu/instrumentation/imacs.
Table 3.2. Fibre+lenslet coupled integral field instruments
Tel. DT Ω dΩ
Instrument Method (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
PMAS Calar Alto 3.5 64. 0.5 256 0.11 9400 1000 0.15
3.5 64. 0.5 256 0.52 1930 1000 0.15
3.5 144. 0.75 256 0.11 9400 1000 0.15
3.5 144. 0.75 256 0.52 1930 1000 0.15
3.5 256. 1.0 256 0.11 9400 1000 0.15
3.5 256. 1.0 256 0.52 1930 1000 0.15
SPIRAL AAT 3.9 251. 0.49 512 0.29 1700 495 0.25
3.9 251. 0.49 512 0.07 7500 495 0.25
MPFS SAO 6.0 256. 1.0 256 0.12 8800 1024 0.045
6.0 64. 0.25 256 0.47 2200 1024 0.045
IMACS-IFU Magellan 6.5 62.0 0.031 2000 0.61 2500 4096 0.19
6.5 37.7 0.031 1200 0.31 7500 2340 0.17
GMOS Gemini 8.0 49.6 0.04 1500 0.21 3450 730 ···
8.0 49.6 0.04 1500 0.32 2300 730 ···
8.0 49.6 0.04 1500 0.82 890 730 ···
8.0 24.8 0.04 750 0.42 3450 1460 ···
8.0 49.6 0.04 1500 0.64 2300 1460 ···
8.0 49.6 0.04 1500 1.00 890 1460 ···
VIMOS VLT 8.0 2916. 0.45 6400 0.6 250 150 ···
8.0 698. 0.11 6400 0.6 250 150 ···
8.0 729. 0.45 1600 0.2 2500 500 ···
8.0 174.5 0.11 1600 0.2 2500 500 ···
ARGUS/IFU VLT 8.0 83.9 0.27 315 0.105 11000 1155 ···
8.0 83.9 0.27 315 0.042 39000 1625 ···
ARGUS VLT 8.0 27.7 0.09 315 0.105 11000 1155 ···
8.0 27.7 0.09 315 0.042 39000 1625 ···
none
COHSI UKIRT 3.8 8.5 0.85 100 0.26 500. 128 ···
SMIRFS UKIRT 3.8 24.2 0.34 72 0.023 5500. 128 ···
CIRPASS Gemini 8.0 54.5 0.13 490 0.41 2500. 1024 ···
8.0 54.5 0.13 490 0.085 12000. 1024 ···
8.0 27.0 0.06 490 0.41 2500. 1024 ···
8.0 27.0 0.06 490 0.085 12000. 1024 ···
none
deflections, the slices are re-aligned end to end as in a long slit, which then feeds a
conventional spectrograph.
The latest incarnation is the so-called ‘Advanced Image Slicer’ (AIS) concept: a three-
element system, introduced and nicely illustrated by Allington-Smith et al. (2004). In
short, the slicer mirrors at the telescope focal plane divide it into strips and have power
to place the telescope pupil on the next slicer element. This is desirable to keep these
elements small and the slicer compact. The second element is an array of pupil mirrors
Table 3.3. Slicer coupled integral field instruments
DT Ω dΩ
ESIa Keck 10.0 22.8 1.28 18 0.95 3500 3325 0.14

10.0 15.0 0.56 27 0.95 5200 4950 0.14
10.0 10.0 0.25 40 0.95 7800 7410 0.14
10.0 8.4 0.09 93 0.95 13000 12350 0.14
WiFeS ANU 2.3 775. 1. 775 1.03 3000 3090 ···

2.3 775. 1. 775 0.44 7000 3090 ···
MUSEa VLT 8.0 3600 0.04 9e4 0.67 3000 2000 0.24
UIST UKIRT 3.8 19.8 0.06 344 0.15 3500 512 ···
PIFS Palomar 5.0 51.8 0.45 115 0.23 550 128 0.22
5.0 51.8 0.45 115 0.10 1300 128 0.22
NIFSa Gemini 8.0 9.0 0.01 900 0.19 5300 1007 ···
GNIRSa Gemini 8.0 15.4 0.023 684 0.301 1700 512 ···
8.0 15.4 0.023 684 0.087 5900 512 ···
SPIFFI VLT 8.0 0.54 0.006 1024 0.34 3000 1024 0.3
8.0 10.2 0.001 1024 0.34 3000 1024 0.3
8.0 64.0 0.06 1024 0.34 3000 1024 0.3
KMOSa VLT 8.0 188.0 0.04 4204 0.28 3600 1000 ···
FISICAa GTC 10.4 72.0 0.53 136 0.79 1300 1024 ···
a
Advanced image slicer design.
(one per slice), which reformat the slices into a pseudo-slit, where they form an image
of the sky. A tertiary field lens (a lenslet for each slice) controls the location of the
pupil stop in the spectrograph. This is critical for efficient use of the spectrograph. All-
mirror designs exist for the NIR (FISICA, Eikenberry, 2004b), taking advantage of lower
scattering at longer wavelengths to machine monolithic elements. Catadioptric designs
exist for the optical (MUSE, Henault et al., 2004). Here the pupil lenses replace pupil
mirrors, which aids the geometric layout of the spectrograph system.
The salient features of image slicers are (i) they are the only IFU mode to preserve
all spatial information. All other coupling modes destroy spatial information within the
sampling element, either by fibre scrambling or pupil imaging (below). (ii) Image slicers
are also the most compact at reformatting the focal plane on to the detector. (iii) They
can be used in cryogenic systems and at long wavelengths where fibres do not transmit
(although lenslet arrays also accomplish this; see next section). There are some disad-
vantages, including (iv) scattered light from the slicing mirrors (diamond-turned optics
cannot be used in the optical), and (v) a lack of reformatting freedom. The latter is
perhaps less of a concern given that the image is being preserved. However, for possible
multi-object modes, particular attention must be paid to the design of the required relay
optics to avoid efficiency losses.
We summarize the existing and planned instruments in Table 3.3. The length of the
list, particularly in the planned instruments, marks a sea change over the last few years
away from fibre + lenslet coupling. While slicers originated for NIR instruments – starting
with the now-defunct MPE-3D (Thatte et al., 1994) – the list of planned optical slicers
is extensive. Existing NIR instruments include PIFS (Palomar Integral Field Spectro-
graph; Murphy et al., 1999) and UIST (UKIRT Imager-Spectrometer; Ramsay Howat
et al., 2006)6 on 4 m-class telescopes; NIFS (Near-Infrared Integral Field Spectrograph;
McGregor et al., 2003), GNIRS (Gemini Near Infrared Spectrograph; Allington-Smith
et al., 2004) and SPIFFI (Spectrometer for Infrared Faint Field Imaging; Eisen-
hauer et al., 2003; Iserlohe et al., 2004), on 8 m-class telescopes. SINFONI7 on VLT
(Bonnet et al., 2004) in particular has shown the power of NIR adaptive optics (AO)
coupled to an image slicer at moderate spectral resolution achieving 20–30% throughput.
Future NIR instruments include KMOS (Sharples et al., 2004) – a multi-object system
discussed below – and FISICA. Below we also discuss three planned NIR instruments for
space.
While the only existing optical instrument is ESI (Echellette Spectrograph and Imager;
Sheinis et al., 2002, 2006), future optical instruments include WiFeS (Wide-Field Spec-
trograph; Dopita et al., 2004), SWIFT (Short Wavelength Integral Field Spectrograph;
Goodsall et al., this workshop) and MUSE (Multi-Unit Spectroscopic Explorer; Bacon
et al., 2004 and references therein). ESI is unique in being the only cross-dispersed IFU
system. While the number of spatial elements is modest, ESI has enormous spectral mul-
tiplex (at medium spectral resolution and good efficiency), the largest of any instrument
planned or in existence.
3.2.9 Direct lenslet coupling

This is the most significant departure in grating-dispersed 3D spectroscopy, and therefore
the most interesting. The basic concept consists of pupil-imaging spectroscopy using
lenslets. The same type of lenslet array used in the fibre + lenslet mode creates a pupil
image from each lenslet, which again is smaller than the size of the lenslet. Here, the
array of pupil-images forms the spectrograph input focal surface, or object; no fibres or
slicers reformat the telescope focal plane into long-slit; the two-dimensional array of pupil-
images is preserved. However, the pupil image does not preserve the spatial information
within the lenslet field. These pupil images are dispersed, and then re-imaged at the
output spectrograph image surface.
Because direct lenslet injection preserves the 2D spatial data format, this type of instru-
ment typically offers more spatial coverage or sampling at the expense of spectral infor-
mation. The extent of the spectrum from each pupil image must be truncated to prevent
overlap between pupil images. From the instrument design perspective, what is gained is
significant: The spectrograph field of view grows linearly with Ω, instead of as Ω2 as it
must in a long-slit spectrograph, where the 2D spatial information must be reformatted
into a 1D slit. Hence, this mode is best suited to instruments with the largest Ω or NΩ .
Lenslet-coupled instruments have excellent spatial fill factors, identical to fibre +
lenslet systems, and comparable to slicers. Because this is achieved with fewer opti-
cal elements and no fibres, there is no information loss via FRD, and overall the system
efficiency can be very high. As with fibre + lenslet coupling, concerns about scattered
light from lenslets apply here too. Unlike fibre-coupled modes, there is no control over
spatial reformatting. The spectra can be well packed on to the detector but, as noted
above, the bandpass must be crafted to prevent overlap for a given spectral dispersion,
6
See www.jach.hawaii.edu/UKIRT/instruments/uist/uist.html for sensitivities.
7
SPIFFI + MACAO; Spectrograph for Integral Field Observations in the Near Infrared =
Spectrograph for Infrared Faint Field Imaging + Multi-Application Curvature Adaptive Optics.
Table 3.4. Lenslet-coupled integral field instruments
DT Ω dΩ
SAURON WHT 4.2 1353 0.88 1577 0.11 1213. 128 0.147
4.2 99 0.07 1577 0.10 1475. 150 0.147
OASIS WHT 4.2 1.92 0.002 1100 0.50 1000. 400 ···
4.2 31.0 0.026 1100 0.50 1000. 400 ···
4.2 180. 0.17 1100 0.50 1000. 400 ···
none
OSIRIS Keck 10.4 1.2 0.02 3000 0.12 3400. 400 ···
10.4 30. 0.10 3000 0.12 3400. 400 ···
10.4 0.3 0.02 1019 0.47 3400. 1600 ···
10.4 7.5 0.10 1019 0.47 3400. 1600 ···
none
i.e. there is limited spectral coverage at a given resolution. Spectral extraction is critical
to minimize crosstalk while maximizing S/N.
Existing optical systems (SAURON, Bacon et al., 2001; OASIS, McDermid et al., 2004)
have relatively low dispersion due to grism limitations, although the grisms allow for very
compact, undeviated systems. Grating-dispersed systems do exists in the NIR (OSIRIS,
Larkin et al., 2003). Future systems with VPH grisms and gratings will have even higher
efficiency; the coupling mode is well suited to articulated camera spectrographs. The
systems summarized in Table 3.4 are designed to exploit superb image quality with
fine spatial sampling (OASIS and OSIRIS are coupled to AO). While they cannot take
advantage of high dispersion without becoming read-noise limited, systems with larger
specific grasp could be optimized for high spectral resolution.
3.2.10 Filtered multi-slit (FMS) coupling

The notion of direct lenslet coupling motivates a poor person’s alternative which returns
the riches of preserving spatial information. The concept is to use a conventional, multi-
object imaging spectrograph with a narrow-band filter, and a slit mask of multi slits in
a grid pattern with grid spacing tailored to the desired dispersion of the grating. This is
illustrated in Figure 3.11. Spatial multiplexing is increased via filtering. While this only
offers sparse spatial sampling, it preserves spatial information (unlike any other mode
except slicing) and can easily be adapted to existing spectrographs.
The notion of filtering to increase spatial multiplex has been used for multi-object spec-
troscopy, e.g. Yee et al. (1996) in the context of redshift surveys using MOS on CFHT (Le
Fevre et al., 1994). Likewise, fibre + lenslet-coupled IFUs, such as VIMOS and GMOS,
use filtering as an option to prevent spectral overlap in configurations with multiple par-
allel pseudo-slits; this is designed to permit trade-offs in spatial versus spectral coverage.
Long slit Multi-slit + NB filter

4096 pixels
8 arcmin slit
3750 pixels
6144 pixels 4’ slitlets

θ 2’ spacing
>
>λ
1’ slitlets
1.4’ spacing θ
>
>θ
δλ
<
7.5’’ slitlets
FMS 17’’ spacing
Figure 3.11. Filtered multi-slit schematic for SALT’s Robert Stobie Spectrograph (RSS). The
five panels on the left left show the progression from long slit, to filtered long slit, to two different
grids of filtered multi slits. Both are tuned to a high dispersion, 10 nm bandpass, and achieve
a spatial multiplex gain of 3 over a pure long slit, with a ×6 loss in bandpass. Higher spatial
multiplex (two to 10 times) is achieved at lower spectral resolution. The RSS Fabry–Perot mode
is shown for reference. The two rightmost panels show an overlay on a nearby, face-on galaxy
and some on-telescope calibration data for that slit-mask.
What is described here is more like the multi-object mode, but instead uses a regular
grid of slitlets. This is well-suited, for example, to observing single, extended sources.
An example of this type of instrument is the SALT Robert Stobie Spectrograph (RSS),
a prime focus imaging spectrograph with an 8 arcmin field of view, articulating camera,
VPH grating suite, dual Fabry–Perot etalons, and R = 50 order-blocking filters (Burgh
et al., 2003; Kobulnicky et al., 2003). The latter can be used with the multi-slit masks
to gain a factor of 3 in spatial multiplex at the highest spectral resolutions (R = 10,000
with a 10 nm bandpass). At lower resolutions (and fixed bandpass), the slit-packing can
be increased by factors of 2 to 10, such that the gain in spatial multiplex is comparable
to the loss of a factor of 5–6 in spectral multiplex in this particular case (the system is
designed for large spectral multiplex). Even at high spectral resolution what is gained
– beyond the spatial multiplex – is the ability to gain 2D spatial mapping in a single
exposure. On balance, what is lost and gained is comparable from a purely information
standpoint, and hence the choice is, as always, science driven. For the study of nearby
galaxy kinematics, this is an outstanding approach.
3.2.11 Multi-object configurations

Multi-object 3D spectroscopy is a major path for future instrumentation, although it
already exists today in one fabulous instrument: FLAMES/GIRAFFE. Here we are talk-
ing about instruments with multiple, independently positionable IFUs. Returning to
our so-called ‘grand’ merit function, it is for just these types of instruments that Ωs is
relevant.
The most obvious way to feed such an instrument is with fibre or fibre + lenslet
bundles (e.g. FLAMES/GIRAFFE). Fibre-based systems provide flexibility for spatial
positioning, but for cryogenic NIR instruments lenslets or slicers may be required. This
necessitates relay optics, which are more mechanically challenging to design and build,
and the introduction of additional surfaces that lead to lowered throughput. Sharples
et al. (2004) have considered the multiple deployable slicer design for KMOS. It is also
possible to implement direct lenslet coupling (pupil imaging), as demonstrated by the
MUSE concept (Henault et al., 2004), albeit in the context of splitting up a monolithic
field into chunks fed to separate spectrographs.
3.2.12 Summary of considerations

The various coupling methods discussed above present different opportunities for down-
selecting information, and packing three into two dimensions in ways that trade quality
versus quantity.
Information selection and reformatting

Fibre + lenslet, slicers and lenslet modes yield comparable spatial, telescope focal-surface
sampling, while pure fibre systems have at best 65% integral coverage. Fibre-based sys-
tems, however, offer the most flexibility in reformatting telescope-to-spectrograph focal
surfaces. Slicers and FMS preserve full spatial information, but only slicers preserve full
integral spatial information. As a result of this coherency, slicers can give the most effi-
cient packing on the detector. In terms of spectral information, lenslets and FMS have
limited sampling, but other coupling modes all essentially feed long-slit spectrographs
and are therefore comparable.
Coverage versus purity

Scattered-light and cross-talk limit signal purity, but to avoid their deleterious effects
requires less efficient use of the detector by, for example, broader spacing of fibres in the
pseudo-slit, or band-limiting filters, thereby limiting coverage in either or both spatial
and spectral dimensions. The trade-off optimization should be science driven. Within this
context, pure fibre systems and FMS minimize scattered light, although fibre azimuthal
scrambling broadens potential cross-talk between spatial channels of the spectrograph.
Slicer systems, again by virtue of the spatial coherency of each slice, are able to utilize
detector real-estate while maintaining signal purity.
Sky subtraction
There are four primary issues concerning, and that are root causes of, sky subtraction
problems in spectroscopy: (i) low dispersion (sky lines contribute overwhelming shot-
noise); (ii) aberrations and non-locality (skyline profiles vary with field angle – spectral
and spatial – and time); (iii) stability (instrument); and (iv) under-sampling (compound
problems of field-dependent aberrations and flexure). All of these conditions are further
compounded if there is fringing on the CCD.
The solutions to these problems are instrumental, observational and algorithmic, i.e.
in the approach to the data analysis. The instrumental solution involves having a well-
sampled, high-resolution and stable system (you get what you pay for). Fibre-based
systems offer the most mapping flexibility, which is critical for spectrographs with
aberration-limited spectral image-quality. Pupil imaging (lenslets with or without fibres)
may offer advantages for HET/SALT-style telescopes, again if sky subtraction is spec-
trograph aberration-limited.
The observational approach includes (i) beam-switching, where object and sky expo-
sures are interleaved, and (ii) nod-and-shuffle, where charge is shuffled on the detector in
concert with telescope nods between object and sky positions. Both of these approaches
have a 50% efficiency in either on-source exposure or in the number of sources that can
be observed (the on-detector source packing fraction).
An algorithmic approach entails aberration modelling, which is well suited to any of the
coupling methods that feed a spectrograph in a pseudo long slit. The question is to what
extent data analysis can compensate for instrumental limitations and avoid inefficient
observational protocol.
Some examples exist of telescope-time-efficient sky-subtraction algorithms; these are
solutions that do not require beam-switching or nod-and-shuffle. For example, Lissandrini
et al. (1994) identify flux and wavelength calibration, as well as scattered light, as the
dominant problems in their fibre-fed spectroscopic data. They use sky lines for second-
order flux calibration (after flat fields), model scattered light from neighbouring fibres
and map image distortions in pixel space to obtain accurate wavelength calibration. The
improvement is dramatic. Bershady et al. (2005) show that higher-order aberrations are
important; wavelength calibration is critical, but so too is line shape. They describe a
recipe for subtracting the continuum and fitting each spectral channel with a low-order
polynomial in the spatial dimension of the datacube. The algorithm works spectacularly
well for sources with narrow line emission with significant spectral channel offsets (e.g.
high internal dispersion as in a rapidly rotating galaxy, or intrinsically large velocity range
as in a redshift survey) and well-sampled data. For other instruments or sources (poor
sampling, low dispersion, broad lines, small velocity range), if aberrations are significant,
more dedicated sky fibres are needed. On balance, the optical stability of the instrument
is critical.
Are these post-facto algorithmic solutions 100% efficient? Not quite. One still needs
to sample sky, but, as derived in Bershady et al. (2004), the fraction of spatial elements
devoted to sky is relatively low (under 10%, and falling below 3% when the number of
spatial resolution elements exceeds 1000). So here is a case where, with a stable spectro-
graph, considerable efficiency may be gained by employing the right processing algorithm.
Consequently, fibre-fed, bench-mounted spectrographs offer the greatest opportunities to
realize these gains. Regardless of spectrograph type and feed, attention to modelling
optical aberrations is critical for good sky subtraction (Kelson, 2003; Viton and Milliard,
2003).
3.3 Interferometry I: Fabry–Perot interferometry

Fabry–Perot interferometry (FPI) provides a powerful tool for 3D spectroscopy because
FPI is field-widening relative to grating-dispersed systems. That is to say, higher spec-
tral resolution can be achieved with FPI for a given instrument beam size and entrance
aperture. This has long been recognized in astronomy. Unfortunately, the breadth of
applications of FPI to sample the datacube has been under-utilized in astronomy. Astro-
nomical applications almost exclusively use FPs as monochromators, i.e. field-dependent,
tunable filters. This allows for a premium on spatial multiplex at the loss of all spectral
multiplex at a given spatial field angle. Multi-order spectral multiplex can be regained
via additional grating dispersion, as noted below; however, in astronomical applications,
this is largely a concept (with one exception). Further, it is also possible to use FPs
for spectroscopy. In this mode, FPI yields the converse trade in spatial versus spectral
multiplex. There is again only one example of such an existing instrument. In short,
FPI to date has offered two (orthogonal) extremes in sampling the datacube. The third
b) λ
grating
λ/λ0 ∼ θ
a)
θ
θ max θ max
∆λ
l n ∆λ θ
FP (or Michelson)
λ/λ0 ∼ θ 2
∆ path = 2 nl cos θ
Figure 3.12. Basic concept of etalon interference (a), and FP versus grating spectral resolution
as a function of angular aperture, θ (b).
dimension (bandpass or field sampling on the sky) has been gained via the temporal
domain, i.e. multiple observations. In this sense FPI has not yet been implemented for
truly 3D spectroscopy.
The basic principles of FPI, in the context of astronomical monochromators, can be
found in Geake et al. (1959), Vaughan (1967) and many other references. We summarize
the salient aspects to highlight here the field-widened capabilities (we are indebted to R.
Reynolds for the structure of the formal development). We discuss and give examples of
the two different FPI applications noted above and sketch how one might balance spatial
and spectral multiplex in future 3D instruments.
3.3.1 Basic concepts and field widening

Etalons (high-precision flat glass plates) are parallel-spaced by some distance l, filled
with gas of refractive index n and coated to have high reflectivity. Light incident at some
angle θ produces internal reflections, with transmission when the added path (∆path =
2nl cos θ) between reflections yields positive interference (Figure 3.12a). The ratio of
transmitted to incident intensity, It /Ii , is given approximately by an Airy function with
peaks (It = Ii ) when ∆path = mλ, where m is the order. Given the geometry, this yields
an angular dependence to the transmitted wavelength: λ = (2nl/m) cos θ. This can be
compared to the grating equation (Littrow configurations for simplicity), where λ =
2nΛg /m sin θ. At small angles, this means that the instrument entrance aperture can
be larger in angle for an FP compared to a grating spectrograph for the same δλ, as
illustrated in Figure 3.12b. In other words, an FP system is field-widened for the same
spectral resolution (see also Roesler, 1974; Thorne, 1988).
The central wavelength of the FP is controlled by tuning the gap (l) or pressure
(index n). The free spectral range is given by the spacing between Airy function peaks in
wavelength: Q = 1/2nl cos θ. Order-blocking filters are needed to suppress other orders.
Double etalons suppress the Lorentian wings in the Airy function. The resolution, which
is the full width at half maximum of the Airy formula peak, is given by: R = λ/δλ =
2l cos θNℜ /λ = mNℜ , where Nℜ is the reflective finesse defined as Nℜ = πℜ1/2 /(1 − ℜ),
and ℜ is the reflectivity. The finesse is equivalent to roughly the number of back and forth
reflections; a typical value is ∼30 (for a more detailed discussion, see Tanaka et al., 1985).
This implies that the spectral resolution, R, is roughly the total path difference divided
by the wavelength. High spectral resolution requires high finesse or high order, with the
gap size tuned for the desired wavelength. This also achieves high contrast between the
maximum and minimum transmittance between orders: Im ax /Im in = (1 + ℜ)2 /(1 − ℜ) =
1 + π42 Nℜ2 . Herbst and Beckwidth (1988) provide a nice illustration of these quantities.
3.3.2 FP monochromators
FPs are conventionally thought of as being used with collimated beams (a mini-review
of such instruments from that era is presented by Bland and Tully, 1989). In this case,
there is the classic radial wavelength dependence in the image plane. At low spectral
resolution the bandpass can be made nearly constant over a large field of view (Jones
et al., 2002), as follows.
One way to characterize an etalon is by the size of its ‘bull’s eye’, or Jacquinot spot
(Jacquinot, 1954). The bull’s eye refers to the physical
' angle θ such that λ0 /|λ0 − λθ | <
R, and is given by θm ax = cos−1 (1 − 1/R) ≃ 2/R. This quantity is independent of
the telescope and is a property of the etalon. By coupling to a telescope, it is possible
to modify the angular scale (α) sampled on the sky by the bull’s eye. Since A × Ω is
conserved, α = θDe /DT , where De is the etalon diameter and DT the telescope diameter.
FPs can, however, be used in converging (or diverging) beams, even near a focus
(Bland-Hawthorn et al., 2001). Some examples include the optical FP on the 3.6 m CFHT,
when used with the AO Bonnette (AOB)8 and the future F2T2, a near-infrared double-
etalon system for FLAMINGOS-2 (8 m Gemini; Eikenberry et al., 2004a; Scott et al.,
2006). Image information is preserved by sampling the beam at a downstream focus, but
the spectral resolution is lowered (for a given finesse) at any spatial location because
each field angle on the sky is mapped into a range of physical angles through the etalon.
The degradation is not particularly severe for lower-finesse etalons or very slow beams.
The FLAMINGOS-2 multi-conjugate adapative optics (MCAO) focus for F2T2 is f /30,
and the AOB F-P beam is f /40. If the total angular field of view is much smaller than
the beam angle, or the focus is made telecentric, the bandpass is constant across field
angles on the sky, and the system forms a highly uniform tunable filter. The AOB optics
are not telecentric; this produces a radial degradation in the resolution.
3.3.3 FP spectrometers
Alternatively, the full spectral information can be extracted at the loss of the spatial
information by placing the etalons at or near a telecentric focus and sampling the pupil
in a collimated beam. The Wisconsin Hα Mapper (WHAM; Reynolds et al., 1998) is
the only astronomical example of this type of instrument. In this instance, the light is
collimated after it passes through the etalons, never refocused, and a detector is placed
at the pupil formed by the collimator. Field position on the detector contains spectral
information: each radius corresponds to a different wavelength. This is similar to the
monochromator application, except in this case each radial location on the detector has
a superposition from all spatial locations on the sky within the instrument entrance
aperture.
3.3.4 3D FP spectrophotometers
Grating-dispersed FPI
Arguably the most interesting FP monochromator mode is to eliminate the order-blocking
filters and grating-disperse the output beam to separate the orders on to the detector to
increase the spectral multiplex. See, for example, le Coarer et al.’s (1995) description of
PYTHEAS. Baldry and Bland-Hawthorn (2000) work out a particularly compelling case
8
For an earlier incarnation on this telescope, see www.cfht.hawaii.edu/Instruments/
Spectroscopy/Fabry-Perot/ and Joncas and Roy (1984).
for a cross-dispersed echelle system. The gain in spectral multiplex does not necessarily
cost spatial multiplex. In practice, some FPs are in spectrographs where they underfill
the detector and usable field in the image plane (e.g. RSS and F2T2). If the dispersion is
significantly greater than the etalon resolution, then, in addition to spectral multiplex,
this mode adds band-limited slitless spectroscopy in each FP order.
Pupil-imaging FPI
The above discussion frames the notion that detection downstream of an etalon at the
pupil of a collimated beam provides spectral information but no spatial information,
while detection at a focal surface provides the complement. A simple ray-trace shows
that between these two locations spectral and spatial information are mixed. By using
pupil imaging at the system input via a lenslet array (Section 3.2.9), detection at an
intermediate surface in a converging beam can separate spatial and spectral information.
Although this has never been done, in principle this could balance spatial and spectral
multiplex and allow for true 3D spectroscopy in future, field-widened instruments.
3.3.5 Sky stability

Because spectral channels are not observed simultaneously in monochromatic modes,
atmospheric changes must be calibrated (in the context of TAURUS, see, for example,
Atherton et al., 1982). Field stars may suffice if they are sufficiently featureless over the
scanned wavelength range. Built-in calibration is desirable, which can be achieved, for
example, via a diachronic feeding a monitoring camera. This capability is designed for a
new generation of instruments (e.g. ARIES, T. Williams, private communication).
3.3.6 Examples of instruments

Two extremes in FP instrumentation are highlighted by the RSS imaging FP (Williams
et al., 2002) and the WHAM non-imaging FP. Both have 150 mm etalons, but the RSS
system is coupled to a 9.2 m telescope with an 8 arcmin field of view, 0.2 arcsec sampling
and spectral resolutions of 500, 1250, 5000, and 12,500. In contrast, WHAM is coupled
to a 0.6 m telescope, with a 1 deg field of view and angular resolution, spectral resolution
of R = 25,000, and spectral coverage of about 166 resolution elements for one spatial
element.
There are a large number of existing FP monochromators (also know as tunable filters),
indicated even by the following incomplete list. Optical systems include, but are not
! PUMA (2.1 m OAN-SPM; Rosado et al., 1995);

limited to:
! RFP (1 m and 4 m CTIO; e.g. Sluit and Williams, 2006);

! CIGALE (3.6 m ESO [European Southern Observatory] and 1.9 m OHP; Boulesteix
! FaNTOmM (1.6 m OMM, 1.9 m OHP, and 3.6 m CFHT; Canada–France–Hawaii

et al., 1984);
! Goddard FP (3.5 m APO; Gelderman et al., 1995);

Telescope; Hernandez et al., 2003);
! SCORPIO FP (6 m SAO; Afanasiev and Moiseev, 2005);

! IMACS FP (Inamori Magellan Cassegrain Spectrograph FP; 6.5 m Magellan; Dressler
et al., 2006);
as well as the above-mentioned CFHT FP etalons, which can be used with the AOB as
well as the MOS (Multi-Object Imaging Spectrograph) and SIS systems. The most widely
cited system is TTF/TAURUS-II (3.9 m AAT; Anglo-Australian Telescope; 4.2 m WHT;
Gordon et al., 2000 and references therein).
! NIC-FPS (3.5 m Arc; Hearty et al., 2004);

Existing infrared instruments include:
! GriF (3.6 m CFHT; Clenet et al., 2002);

! PUMILA (2.1 m OAN-SPM; Rosado et al., 1998);
! UFTI (3.8 m UKIRT; United Kingdom Infrared Telescope; Roche et al., 2003); and
! NaCo (8 m VLT; Hartung et al., 2004; Iserlohe et al., 2004).
GriF, NaCo and F2T2 are AO-fed. By virtue of their use in collimated beams, many of
the FP systems are designed to be transportable between instruments (i.e. spectrographs
or focal reducers) and telescopes. Future instruments include the optical OSIRIS (Optical
System for Imaging and Low/intermediate-Resolution Integrated Spectroscopy; 10.4 m
GTC; Gran Telescopio CANARIAS) and near-infrared FGS-TF (6.5 m JWST; James
Webb Space Telescope; Davila et al., 2004) and F2T2 (above). These systems span a
wide range of wavelength, spectral and spatial resolution. One attribute they have in
common is a spectral multiplex of unity.
3.4 Interferometry II: spatial heterodyne spectroscopy

A spatial heterodyne spectrometer (SHS) is a Michelson interferometer with gratings
replacing the mirrors. The principles of operation are described and illustrated by Har-
lander et al. (1992); this is a paper that is well worth careful study.9 Briefly, each grating
diffracts light at wavelength-dependent angles. Because of the 90 degree fold between the
two beams, the wavefronts at a given wavelength are tilted with respect to each other
after beam recombination. This tilting produces a sinusoidal interference pattern with a
frequency dependent on the tilt angle. The degree of tilt is a function of wavelength, sim-
ply due to the grating diffraction, and hence the interference pattern frequency records
the wavelength information.
It is easiest to conceptualize this in terms of two identical gratings (as illustrated
by Harlander et al., 1992 in their figures 2 and 3), but in principle the gratings do
not need to be the same. Wavefronts produce interference patterns with frequencies
set by wavelength, with the central wavelength producing no interference. Hence, the
signal is heterodyned about the frequency of the central wavelength. Resolution is set
by the grating aperture diameter because this sets the wavelength (i.e. angular tilt)
that minimally departs from the central wavelength which can produce the first (lowest)
frequency for interference. Bandwidth is set by the length of the detector, i.e. how many
frequencies can be sampled depends on the number of pixels.
The advantage of an SHS over a Michelson is that no stepping is required to gain the
full spectral information, but the field of view is reduced. The SHS can be fed with a
long slit or lenslet array, although with the latter a band-limiting filter is needed (as with
a conventional dispersed spectrograph). As with a Michelson, however, field-widening
is possible via prisms. In the SHS application, the prisms give gratings the geometric
appearance of being more perpendicular to the optical axis; hence, larger field angles are
mapped within the beam deviation producing the lowest-order interference fringe. Cross-
dispersion is possible (by tilting one of the gratings about the optical axis), but the same
fundamental limits apply concerning 3D information formatted into a 2D detector.
One of the problems with the standard Michelson or SHS interferometer is that their
geometry throws out half the light right from the start. Non-lossy geometries are possible.
Harlander et al. (1992) give an example of working off-axis on the collimating mirror (see
their figure 5). This is a perfect application for holographic gratings. Transmission-grating
9
The presentation here benefited from discussion with J. Harlander, A. Sheinis, R. Reynolds,
F. Roesler and E. Merkowitz.
a) b)
Figure 3.13. Total and specific grasp versus spectral power for a range of instruments on 4-m-
and 10-m-class telescopes (solid and dashed lines, respectively) partially updated from Bershady
et al. (2005). See text for comments on instrument efficiency.
geometries would eliminate the need to go off-axis and probably allow for larger field.
Another approach is a Mach–Zender-style interferometer (Douglas et al., 1990). The
latter requires twice the detector real estate for the same number of spectral resolution
elements.
The primary advantage of an SHS is that it allows for very high spectral resolution for
a given solid angle relative to a conventional, grating-dispersed spectrograph. The SHS
is field-widened like an FP. This means the SHS can be built for low cost even on large
telescopes because the optics are small.
However, because the signal is in the form of an interferogram, there is what is known
as the ‘multiplex disadvantage.’ This can be expressed as the S/N performance of the
SHS relative to a grating spectrograph: S/NSHS = S/NGS (f /2)1/2 (SSHS /SGS )1/2 , where
S/NSHS and S/NGS are the signal to noise in SHS and grating spectrometer, respectively,
SSHS and SGS are the total photon signal, respectively, and f is the fraction of total signal
in a given spectral channel (f < 1, and decreases with bandwidth). In words, this means
that an SHS loses competitiveness with grating-dispersed spectrographs when the band-
pass is large. This has implications for design and use. Clearly, one must make SSHS and
f as large as possible. The small compact optics of an SHS system lend themselves to effi-
ciency optimization. To make f as large as possible, one must choose a small bandwidth
(but more than a Fabry–Perot monochromator) and remove OH lines via pre-filtering, or
by selecting bandpasses between them. Returning to Figure 3.1, SHS is between an FP
monochromator and other IFS methods and will therefore have application to a broad
range of science programmes that seek high spectral resolution over a limited bandpass
with good spatial coverage.
3.5 Summary of existing instruments

Here, we explore the sampled parameter space in spatial versus spectral information, as
well as coverage versus resolution, starting with grasp and spectral power (Figure 3.13).
Recall that, because reliable, consistent measurements of efficiency are unavailable for
most instruments, we use grasp instead of etendue (warning: we really want etendue).
Note, however, that there is a factor of 6 range in the known efficiencies of instruments
tabulated in this chapter. Further, note that there are two ways of viewing the specific
grasp. From the perspective of staying photon-limited at high spectral resolution, high
specific grasp is important. The ‘flip side’ is that low specific grasp implies high angular
resolution.
Figure 3.14. Spatial resolution (a) and specific grasp (b) versus spectral power for all instru-
ments in Tables 3.1–3.4, highlighting differences between optical (filled symbols) and NIR (open
symbols), as well as between different coupling methods (labelled).
Figure 3.15. Total grasp versus spectral power (a) and the number of spatial (NΩ ) versus
spectral (NR ) resolution elements for all instruments in Tables 3.1–3.4, highlighting differences
between optical (filled symbols) and NIR (open symbols). Dashed lines are at constant (labelled)
total information (NR × NΩ ).
Figure 3.14 shows that spatial resolution is higher in NIR instruments, while spectral
resolution is higher in optical instruments. Fibre IFUs have the largest specific grasp,
reflected in the bifurcation seen in spatial resolution, i.e. fibre-fed instruments have large
footprints per element (dΩ). There is a trend of decreasing specific grasp going from fibre
+lenslet, lenslet, and finally to slicers. ESI has unusually large A × dΩ for a slicer; RSS
in FMS mode has the highest specific grasp overall.
Figures 3.14 and 3.15 together show that optical and near-infrared instruments trade
spatial resolution for grasp; there are no high-grasp NIR instruments; the highest spectral
power instruments are optical. Optical and near-infrared instruments sample comparable
total information, with optical instruments sampling a broader range of trades between
spatial versus spectral information. Older NIR instruments clearly suffer from being
detector-size limited. IMACS-IFU stands out as having a significantly larger number of
total information elements, NR × NΩ , and in this sense is on par with future-generation
instruments.
3.6 The extended-source domain

One area of extragalactic science is clearly under-sampled by existing instrumentation,
namely high spectral resolution yet low surface brightness 3D spectroscopy of extended
sources. The scientific impetus is for detailed nebular studies (ionization, density, metal-
licity, abundances) of not only compact HII regions, but to extend such study to the
diffuse, ionized gas. Likewise, a significant fraction of the stellar light in galaxies is in
extended distributions at low surface-brightness, i.e. below the night sky background. The
kinematic and chemical properties of these stars is largely unknown outside of resolved
populations in the Local Group. Stellar kinematics of galaxies on spatially resolved scales
are required to dissect the mass distribution and detailed dynamics of disc, bulge and
halo components. This information is effectively the Rosetta Stone for deciphering how
galaxies have assembled.
One concern with most existing IFU spectrographs is their focus on very fine spa-
tial sampling. Referring back to Figure 3.4, on telescopes as small as only 10 m, this
severely limits the spectral resolution that can be achieved at sub-arcsec sampling in the
photon-limited regime. For example, FLAMES/GIRAFFE is unusual in its high spectral
resolutions of 10–40,000. Each IFU unit is 2 × 3 arcsec with 20 rectangular microlenses
sampling only 0.52 × 0.52 arcsec; this is equivalent to a 1 arcsec fibre on 3.5 m telescope.
The instrument is very close to the photon detector-limited divide. The IMACS-IFU
should be in a similar domain at its high spectral-resolution limit.
There is no question that FLAMES/GIRAFFE has proven spectacular for emission
line work, particularly if line emission is clumpy and unresolved, e.g. ionized gas kine-
matics of distant galaxies (Flores et al., 2004). The need for high angular resolution
in the distant-galaxy kinematic game is paramount. Even with ∼0.5 arcsec resolution,
Hubble Space Telescope images are needed to super-resolve the IFU data (Flores et al.,
2004). It will be difficult, however, to use this same facility to study diffuse gas or the
stellar continuum in resolved sources. Furthermore, resolved structures at high redshift
are all at apparently low surface brightness because of cosmological dimming. To stay
photon-limited, observing in the low surface brightness regime requires lower spectral
resolution, larger apertures (dΩ) or larger telescopes. Will this be addressed by future
instrumentation?
3.7 Future instruments

The next generation of instruments will compete on both space-based platforms such as
the James Webb Space Telescope (JWST) and ground-based telescopes reaching 30 m or
more in diameter. Why build these bigger telescopes? The argument of simply collecting
more photons is compelling but not sufficient. New facilities, which come at increas-
ingly greater cost, must yield gains above the linear increase in area. Such ‘windfalls’
may include overcoming detector noise, the diffraction limit (at long wavelengths), back-
grounds (in the case of space-borne platforms) or critical combinations thereof.
A discussion of backgrounds and the relative merits and niches of 8 m-class space-borne
telescopes such as JWST, and large 30 m-class ground-based telescopes was vetted in the
early planning stages of what was once known as ‘MAXimum Aperture Telescopes’,
or MAXAT. Gillett and Mountain (1998)10 pointed out that a cooled spacecraft has
significantly lower background in the infrared compared to the ground, even at high
spectral resolution. This contrast is dramatic for λ > 2.5 µm, i.e., in the thermal IR.
However, they calculated that above R ∼ 1000, 8 m-class space telescopes are detector-
limited at any wavelength, assuming 0.05 arcsec apertures, a generous system through-
put and realistic detectors. They constructed a competitiveness criteria that assumed
diffraction-limited performance for stellar imaging or spectroscopy. Compared to JWST,
they concluded that ground-based telescopes can be competitive at λ < 2.5 µm for imag-
ing if DT > 20 m, and for diffraction-limited spectroscopy at R > 1000 for DT > 8 m at
any wavelength. These considerations have been influential in the planning and design
of future-generation instruments in the era of JWST.
3.7.1 Ground-based instruments on 10 m telescopes

Given rapid growth in 3D spectroscopy, we expect many new and retro-fitted systems in
the coming years. We outline three instruments – MUSE, VIRUS and KMOS – because
they highlight the common themes of object and instrument multiplexing. (US scientists
will note that two of these systems are on the VLT.) Object multiplexing is a departure for
3D instrumentation; instrument multiplexing is a departure overall. The basic parameters
of these three instruments are summarized in Tables 3.1 and 3.3.
Both MUSE and VIRUS offer unprecedented spatial sampling. MUSE provides a truly
integral 1 arcmin2 area, sampled at the 0.2 arcsec scale, accomplished via image slicing
(AIS-type). The most significant portions of the system are the slicers, which must per-
form well (with little scattered light) in the optical, and the field-partitioning between
a bank of 24 identical spectrographs. In comparison, the individual spectrographs are
modest, albeit high-efficiency, articulated VPH-grating systems.
VIRUS uses the same notion of a replicated spectrograph unit (also articulated VPH-
grating systems), but in this case fed by bare fibres on a much coarser scale (25× larger
dΩ). The field sampling is sparse; hence, this instrument follows directly in the path of
bare-fibre IFUs. What stands out in the VIRUS design is the ‘ultra-cheap’ notion of the
spectrograph unit, i.e. by building many replicated units, costs are lowered by economies
of scale. Such a demonstration has important implications for future large-telescope
instrument design (how large must the replication scale be to manifest significant econ-
omy?). If the full replication of 132 spectrographs is accomplished, this will be by far the
widest-field (largest grasp) IFU in existence, probably for years to come. To achieve the
VIRUS-132 goal requires an all-new wide-field prime-focus spherical-aberration corrector
yielding a 16 arcmin science-grade field for the HET; this is a significant optomechanical
challenge in itself.
MUSE, in contrast to VIRUS, has 9× less total grasp, but almost 3× more spatial
elements (NΩ ). In other words, both stand out as remarkable in spatial sampling in
their own way. The differences in spatial resolution versus coverage between MUSE and
VIRUS lies in their respective science themes. VIRUS is designed as a precision-cosmology
engine to measure the baryon oscillations by detecting z ∼ 3 Lyα-emitters, and using their
distribution as a density tracer. These sources are relatively rare (in surface-density to
a given detected flux), although the exact flux density relation is still uncertain. Rather
low spectral resolution (R < 1000) is needed, since only line identification (in the blue
where backgrounds are low) and redshifts are required.
10
See also the AURA MAXAT Final Report (1999), www.gemini.edu/science/maxat.
Table 3.5. Future TMT integral field instruments
DT Ω dΩ
Instrument Coupling (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
IRMOS slicer 30. 40. 0.01 4000 0.25 2000 500 ···
IRMOS slicer 30. 40. 0.01 4000 0.25 10000 2500 ···
IRIS slicer 30. 0.26 1.6e-5 16384 0.05 4000 200 ···
IRIS slicer 30. 1.33 8.1e-5 16384 0.05 4000 200 ···
IRIS slicer 30. 7.93 4.8e-4 16384 0.05 4000 200 ···
IRIS slicer 30. 41.0 2.4e-4 16384 0.05 4000 200 ···
WFOS fibre+lens 30. 810. 0.56 1440 1.37 5000 6850 0.3
MUSE, in contrast, is designed to probe the detailed internal properties (dynamics,

stellar populations) of galaxy populations over a wide range in redshift and in a rep-
resentative cosmological volume. The aim of this instrument is essentially to enable
spectroscopic versions of many ‘Hubble Deep Fields’, each with sufficient spectral and
spatial information to extract kinematics and line-diagnostics of many thousands of
z < 1 galaxies.
KMOS has much the same science goals as MUSE, with the key distinction of pushing
to higher redshifts by using the NIR to capture the optical rest-frame. By pushing to
higher redshift to gain temporal leverage on the galaxy formation and evolution pro-
cess the source distribution becomes apparently fainter and the NIR backgrounds are
higher. Consequently, on the same size telescope, one is forced to look at intrinsically
more luminous and hence rarer objects. Therefore, the KMOS design moves away from
the notion of a monolithic integral-field, to a 24-probe system in a large, 7.5 arcmin diam-
eter patrol field. Each probe spans a 2.8 × 2.8 arcsec2 area sampled at 0.2 × 0.2 arcsec2 .
While a multi-IFU instrument already exists (again on the VLT) in the optical with
FLAMES/GIRAFFE, the extension to the NIR using slicers with twice the number of
probes will be a significant technical achievement.
What is missing from this suite of remarkable instruments is a design which pushes
forward a significant increase in spectral sampling (spectral power) or specific grasp.
For example, none of these instruments offers over R = 4000 and NR = 2000. This
means, for example, that advances in the study of low surface brightness, dynami-
cally cold (σ < 80 km s−1 ) systems or nebular regions will require additional instrument
innovation.
3.7.2 Ground-based instruments on 30–50 m-class telescopes

We summarize some of the specific examples of TMT 3D spectroscopic instrumentation
in Table 3.5, based on D. Crampton’s overview (Crampton and Simard, 2006).11 TMT
instrument design is largely driven by AO capabilities, where the salient point is that
there are many ‘flavours’ of AO, with associated levels of difficulty and risk (inversely
proportional to their performance in either image quality or field of view, or both).
The IFU-capable TMT instruments include, in order of decreasing AO requirements:
(i) IRMOS, a NIR multi-object integral-field spectrograph fed by the multi-object adap-
tive object system (MOAO), capable of 20 positional, 5 arcsec compensated patches
within a 5 arcmin patrol field; (ii) IRIS, a NIR imager and integral field spectrograph
working at the diffraction limit, fed by the narrow-field facility AO system (NFIRAOS);
11
See also www.tmt.org/observatory/index.html.
Table 3.6. Future space-based integral field instruments
DT Ω dΩ
Instrument Coupling Tel. (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
FGS-TF FP JWST 6.5 38088. 0.018 2.10e7 0.01 100 1 ···

NIRSpec AIS JWST 6.5 9. 0.0056 1600 0.34 3000 1024 · · ·
MIRI AIS JWST 6.5 51.8 0.30 173 1.48 2800 4096 · · ·
SNAP-IFU AIS SNAP 2. 9.0 0.022 400 1.95 100 195 0.44
and (iii) WFOS, an optical wide-field, seeing-limited spectrograph with potential for a
modest-grasp IFU with good spectral power and spectral resolution (R < 6000).
With the exception of WFOS, instrument design is driven by AO considerations
because of the enormous physical size of the image (which scales with mirror diameter for
a constant f -ratio). WFOS is necessarily a monster. The AO-driven focus is suitable for
scientific studies of un- or under-resolved sources (stars, planets, sub-kpc scales in distant
galaxies), and excellent science cases have been developed. Of this excellence there is no
doubt. Of concern is that once wedded to the notion of a very large telescope with no
clear path to building affordable, comparably monstrous instruments, one is forced down
a path, ab initio, of considering only science enabled by high angular resolution. It is not
surprising to note that WFOS – the one non-AO corrected instrument – stands out as
also the one TMT instrument concept that breaks into the high specific-grasp domain at
modest spectral-resolution domain (Figure 3.16). Indeed, as seen in Figure 3.17, WFOS
breaks new ground in terms of its total grasp with a higher spectral power than any
existing instrument (save ESI). To optimize low surface brightness studies, other paths
will need to be forged to push to higher spectral power and resolution at comparably
high grasp.
These same trends are also being played out for instrument design for extremely large
telescopes (ELTs) (e.g. Eisenhauer et al., 2000; Russell et al., 2004). We have focused
on TMT because of the more mature stage of this telescope’s planning. No doubt ELT’s
complement of instruments will open up exciting new capabilities, as demonstrated by the
superb forefront instrumentation on the VLT. The TMT instrumentation programme,
like that of the European ELT, is evolving rapidly. What is presented here is a snapshot
circa late 2005.
3.7.3 Space-borne instruments

We summarize the planned 3D spectroscopic instruments for JWST and SNAP (Super
Nova Acceleration Probe; Aldering et al., 2002) in Table 3.6. (There are other missions,
which include IFUs, also in the planning stages.) On JWST, in remarkable contrast
to HST, three of the four instruments have 3D spectroscopic modes in the near- and
mid-infrared: (i) FGS-TF (of which F2T2 is the ground-based analogue) delivers a 2.3 ×
2.3 arcmin2 field at R ∼ 100, with two cameras covering 1.2 to 4.8 µm; (ii) Near-Infrared
Spectrograph (NIRSpec) (Prieto et al., 2004) has a 3 × 3 arcsec2 IFU using and AIS with
40 3 × 0.075 arcsec2 slices, covering 0.8–5 µm at R = 3000; (iii) Mid-Infrared Instrument
(MIRI) (Wright et al., 2004) has 4 simultaneous image-slicers at R ∼ 3000 feeding four
wave-bands between 5 and 28 µm. Each samples 4.6 × 5.5 arcsec2 (increasing by a factor
of two between the bluest and the reddest channel) with a 0.37 arcsec slit width (changing
by a factor of 4 between the bluest and the reddest channel). Quoted numbers represent
mean values over all channels.
Figure 3.16. Spatial resolution (a) and specific grasp (b) versus spectral power including future
instruments in Tables 3.5 and 3.6: existing (filled circles), future ground-based 10 m-class tele-
scopes (open circles), future TMT (open squares), and future space-borne (open triangles).
The SNAP IFU (Ealet et al., 2003) is designed to identify SNe type out to z ∼ 1.7. As
such, it is unique in being a dual optical–NIR system (0.35–1.7 microns), with a 3 × 3
field using AIS, but very low spectral resolution (R = 100). With its very high expected
efficiency, co-added datasets should yield superb, spatially resolved spectrophotometry
of galaxies on 1–2 kpc scales.
Overall, future space-borne capabilities can be characterized as having 3 × 3 arcsec2
fields mapped with AIS technology with 0.15 arcsec sampling, lower spatial resolution
than TMT. Spectral resolution is in the 100 < R < 3000 range, again lower than TMT.
This is consistent with their being competitive in performance relative to TMT-class
instruments, given Gillett and Mountain’s (1998) argument. However, there are no large-
grasp systems that take full advantage of the low backgrounds of space. There are no
high- or even medium-resolution spectrographs to couple, competitively, to such large
angular apertures. Nonetheless, barring past fiascos, the space-borne missions offer the
guarantee of superlative image quality and low backgrounds extending into the mid-IR,
while ground-based observatories face the intense challenge of developing advanced AO
systems.
3.7.4 Summary of future instruments

While Figure 3.16 shows some of the areas not accessed by currently planned future
instrumentation, at the same time clearly great strides are planned for accessing new
domains in spatial resolution from the ground. This is encouraging because only with
the largest apertures can we stay photon-limited at moderate spectral resolution. JWST
instruments present the unique ability to work at more modest spectral resolution and
still remain source-photon limited. Space-borne instruments, overall, will also provide the
most-stable and best-characterized point-spread functions, a premium for high angular-
resolution spectrophotometry.
Figure 3.17 reveals where new instruments open up new territory: both in added grasp
at high spectral power and simply in more resolution elements (NR × NΩ ). Of particu-
lar note is the thrust towards instruments with many thousands of spectral resolution
elements. These gains are seen for both ground-based and space-borne instruments, on
both 10 m- and 30 m-class telescopes.
Figure 3.17. Total grasp versus spectral power (a) and the number of spatial (NΩ ) versus
spectral (NR ) resolution element (b) including future instruments in Tables 3.5 and 3.6; symbols
as in previous figure.
These gains are made with conceptually conventional grating-dispersed systems or

FP monochromators. Clearly, there is opportunity for less conventional, field-widened
instruments (such as the interferometric concepts discussed above), which can amplify
both grasp and spectral power or spectral resolution. Given the relative novelty of these
approaches, they present higher risk, but potentially higher return, and are best suited
for ground-based development.
Acknowledgements
We would like to thank the Instituto de Astrofı́sica de Canarias (IAC) and Winter School
organizers, the University of Toronto Department of Astronomy and Astrophysics for
their gracious hospitality during a sabbatical year where this work was done, and the
NSF for their financial support of this research (AST-0307417 and AST-0607516).
REFERENCES
Afanasiev V.L., Moiseev, A.V. (2005), Astr. Lett., 31, 193
Afanasiev, V.L, Dodonov, S.N., Sil’chenko, O.K., Vlasyuk, V.V. (1990), SAO preprint,
N54
Aldering, G., Akerlof, C.W., Amanullah, R. ET AL. (2002), Proc. SPIE, 4835, 146
Alighieri, S.S. (2005), in the Proceedings of the ESO Workshop on Science Perspectives for
3D Spectroscopy, (Garching, Germany)
Allende Prieto, C., Majewski, S.R., Schiavon, R. ET AL. (2008) AN, 329, 1081
Allington-Smith, J., Content, R. (1998), PASP, 110, 1216
Allington-Smith, J., Dubbeldam, C.M., Content, R. ET AL. (2004), Proc. SPIE, 5492,
701
Arribas, S., Mediavilla, E., Rasilla, J.L. (1991), ApJ, 369, 260
Arribas, S., del Burgo, C., Carter, D. ET AL. (1998), ASPCS, 152, 149
Atherton, P.D., Taylor, K., Pike, C.D., Harmer, C.F.W., Parker, N.M., Hook, R.N.
(1982), 201, 661
Avila, G., Guinouard, I., Jocou, L., Guillon, F., Balsamo, F. (2003), Proc. SPIE, 4841,
997
Bacon, R., Bauer, S.-M., Bower, R. ET AL. (2004), Proc. SPIE, 5492, 1145
Baldry, I.K., Bland-Hawthorn, J. (2000), PASP, 112, 1112
Baldry, I.K., Bland-Hawthorn, J., Robertson, J.G. (2004), PASP, 116, 403
Baranne, A. (1972), in ESO/CERN Conference on Auxiliary Instrumentation for Large Tele-
scopes, ed. S. Lautse & A. Reiz (Geneva), 227
Barden, S.C., Wade (1988), ASPCS, 3, 113
Barden, S.C., Elston, R., Armandroff, T., Pryor, C. (1993), ASPCS, 37, 223
Barden, S.C., Sawyer, D.G., Honneycutt, R.K. (1998), Proc. SPIE, 3355, 892
Barden, S.C., Arns, J.A., Colburn, W.S., Williams, J.B. (2000), PASP, 112, 809
Bershady, M.A., Andersen, D.R., Harker, J., Ramsey, L.W., Verheijenn, M.A.W.
(2004), PASP, 116, 565
Bershady, M.A., Andersen, D.R., Verheijen, M.A.W., Westfall, K.B., Crawford,
S.M., Swaters, R.A. (2005), ApJS, 156, 311
Bershady, M., Barden, S., Blancho, P.-A. ET AL. (2008) Proc. SPIE, 7014, 70140H-1
Blais-Ouellette, S., Guzman, D., Elgamil, A., Rallison, R. (2004), Proc. SPIE, 5494,
278
Blais-Ouellette, S., Daigle, O., Taylor, K. (2006), Proc. SPIE, 6269, 174
Bland, J., Tully, R.B. (1989), AJ, 98, 723
Bland-Hawthorn, J., van Breugel, W., Gillingham, P.R., Baldry, I.K. (2001), ApJ,
563, 611
Bonnet, H. ET AL. (2004), Msngr, 117, 17
Boulesteix, J., Georgelin, Y., Marcelin, M., Monnet, G. (1984), Proc. SPIE, 445, 37
Burgh, E.B., Nordsieck, K.H., Kobulnicky, H.A., Williams, T.B., O’Donoghue, D.,
Smith, M.P., Percival, J.W. (2003), Proc. SPIE, 4841, 1463
Burgh, E.B., Bershady, M.A., Westfall, K.B., Nordsieck, K.H. (2007), PASP, 119, 1069
Carrasco, E., Parry, I.R. (1994), MNRAS, 271, 1
Clenet, Y., Le Coarer, E., Joncas, G. ET AL. (2002), PASP, 114, 563
Crampton, D., Simard, L. (2006), Proc. SPIE, 6269, 59
Davila, P.S., Bos, B.J., Contreras, J. ET AL. (2004), Proc. SPIE, 5487, 611
Dopita, M.A., Waldron, L.E., McGregor, P. ET AL. (2004), Proc. SPIE, 5492, 262
Douglas, N.G., Butcher, H.R., Melis, M.A. (1990), Ap&SS, 171, 307
Dressler, A., Hare, T., Bigelow, B.C., Osip, D.J. (2006), Proc. SPIE, 6269, 13
Ealet, A., Prieto, E., Bonissent, A. ET AL. (2003), Proc. SPIE, 4850, 1169
Eikenberry, S.S., Elston, R., Raines, S.N. ET AL. (2004a), Proc. SPIE, 5492, 1196
Eikenberry, S.S., Elston, R., Guzman, R. ET AL. (2004b), Proc. SPIE, 5492, 1264
Eisenhauer, F., Tecza, M., Thatte, N., Mengel, S., Hofmann, R., Genzel, R. (2000),
Proceedings of the Backaskog Workshop on Extremely Large Telescopes, ed. T. Andersen,
A. Ardeberg, R. Gilmozzi (Lund/ESO), 57, 292
Eisenhauer, F., Abuter, R., Bickert, K. ET AL. (2003), Proc. SPIE, 4841, 1548
Fabricant, D.G., Hertz, E.N., Szentgyorgyi, A.H. ET AL. (1998), Proc. SPIE, 3355, 285
Fabricant, D.G., Fata, R., Roll, J. ET AL. (2005), PASP, 117, 1411
Flores, H., Peuch, M., Hammer, F., Garrido, O., Hernandez, O. (2004), A&A, 420,
L31
Geake, J.E., Ring, J., Woolf, N.J. (1959), MNRAS, 119, 161
Gelderman, R., Woodgate, B.E., Brown, L.W. (1995), ASPC, 71, 89
Gillett, F., Mountain, M. (1998), ASPC, 133, 42
Gordon, S., Koribalski, B., Houghton, S., Jones, K. (2000), MNRAS, 315, 248
Guerin, J., Felenbok, P. (1988), ASPCS, 3, 52
Hanuschik, R.W. (2003), A&A, 407, 1157
Harlander, J., Reynolds, R.J., Roseler, F.L. (1992), ApJ, 730
Hartung, M., Lidman, C., Ageorges, N., Marco, O., Kasper, M., Clenet, Y. (2004),
Proc. SPIE, 5492, 1531
Haynes, R., Lee, D., Allington-Smith, J. ET AL. (1999), PASP, 111, 1451
Hearty, F.R., Morse, J., Beland, S. ET AL. (2004), Proc. SPIE, 5492, 1623
Henault, F.R., Bacon, R., Content, R. ET AL. (2004), Proc. SPIE, 5249, 134
Herbst, T.M., Beckwidth, S. (1988), PASP, 100, 635
Hernandez, O., Gach, J.-L., Carignan, C., Boulesteix, J. (2003), Proc. SPIE, 4841,
1472
Hill, G.J., MacQueen, P.J., Tejada, C. ET AL. (2004), Proc. SPIE, 5492, 251
Iserlohe, C., Tecza, M., Eisenhauer, F. ET AL. (2004), Proc. SPIE, 5492, 1123
Jacquinot, P. (1954), JOSAA, 44, 761
Joncas, G., Roy, J.R. (1984), PASP, 96, 263
Jones, D.H., Shopbell, P.L., Bland-Hawthorn, J. (2002), MNRAS, 329, 759
Kelson, D. (2003), PASP, 115, 688
Kelz, A., Verheijen, M.A.W., Roth, M.M. ET AL. (2006), PASP, 118, 119
Kenworthy, M.A., Parry, I.R., Ennico, K.A. (1998), ASPCS, 152, 300
Kenworthy, M.A., Parry, I.R., Taylor, K. (2001), PASP, 113, 215
Kobulnicky, H.A., Norsieck, K.H., Burgh, E.B., Smith, M.P., Percival, J.W.,
Williams, T.B., O’Donoghue, D. (2003), Proc. SPIE, 4841, 1634
Koo, D.C., Bershady, M.A., Wirth, G.D., Stanford, S.A., Majewski, S.R. (1994), ApJ,
427, L9
Larkin, J.E., Quirrenbach, A., Krabbe, A. ET AL. (2003), Proc. SPIE, 4841, 1600
Le Coarer, E., Bensammar, S., Comte, G., Gach, J.L., Georgelin, Y. (1995), A&AS,
111, 359
Le Fevre, O., Crampton, D., Felenbok, P., Monnet, G. (1994), A&A, 282, 325
Le Fevre, O., Saisse, M., Mancini, D. ET AL. (2003), Proc. SPIE, 4841, 1671
Lissandrini, C., Cristiani, S., La Franca, F. (1994), PASP, 106, 1157
Lorenzetti, D., Cortecchia, F., D’Alessio, F. ET AL. (2003), Proc. SPIE, 4841, 94
Maihara, T., Iwamuro, F., Yamashita, T. ET AL. (1993), PASP, 105, 940
McDermid, R., Bacon, R., Adam, G., Benn, C., Cappellari, M. (2004), Proc. SPIE, 5492,
822
McGregor, P.J., Hart, J., Conroy, P.G. ET AL. (2003), Proc. SPIE, 4841, 1581
Murphy, T.W., Matthews, K., Soifer, B.T. (1999), PASP, 111, 1176
Nelson, G.W. (1988), ASPCS, 3, 2
Oliveira, A.C., de Oliveira, L.S., dos Santos, J.B. (2005), MNRAS, 356, 1079
Parry, I., Bunker, A., Dean, A. ET AL. (2004), Proc. SPIE, 5492, 1135
Prieto, E., Ferruit, P., Cuby, J.-G., Blanc, P.-E., Le Fevre, O. (2004), Proc. SPIE,
5487, 777
Rallison, R.D., Schicker, S.R. (1992), Proc. SPIE, 1667, 266
Ramsay Howat, S.K., Todd, S., Wells, M., Hastings, P. (2006), New Astronomy Reviews,
50, 3513
Ramsey, L.W. (1988), ASPCS, 3, 26
Ren, D. & Allington-Smith, J. (2002), PASP, 114, 866
Reynolds, R.J., Tufte, S.L., Haffner, L.M., Jaehnig, K., Percival, J.W. (1998), PASA,
15, 14
Roche, P.F., Lucas, P.W., Mackay, C.D. ET AL. (2003), SPIE, 4841, 901
Roesler, F.L. (1974), Methods in Experimental Physics, 12A, Academic Press (San Diego),
Chapter 12
Rosado, M., Langarica, R., Bernal, A. ET AL. (1995), RMxAC, 3, 263
Rosado, M., Cruz-Gonzales, I., Salas, L. ET AL. (1998), SPIE, 3354, 1111
Roth, M.M., Kelz, A., Fechner, T. ET AL. (2005), PASP, 117, 620
Russell, A.P.G., Monnet, G., Quirrenbach, A. ET AL. (2004), Proc. SPIE, 5492, 1796
Saunders, W., Bridges, T., Gillingham, P. ET AL. (2004), Proc. SPIE, 5492, 389
Schmoll, J., Roth, M.M., Laux, U. (2003), PASP, 115, 854
Schmoll, J., Dodsworth, G.N., Content, R., Allington-Smith, J.R. (2004), Proc. SPIE,
5492, 624
Schroeder, D. (2000), Astronomical Optics, Academic Press (San Diego), 2nd Edition
Scott, A., Javed, M., Abraham, R. ET AL. (2006), Proc. SPIE, 6269, 176
Sharples, R.M., Bender, R., Lehnert, M.D. ET AL. (2004), Proc. SPIE, 5492, 1179
Sheinis, A.I., Bolte, M., Epps, H.W. ET AL. (2002), PASP, 114, 851
Sheinis, A.I. ET AL. (2006), Proc. SPIE, (astro-ph/0606176)
Sluit, A.P.N., Williams, T.B. (2006), AJ, 131, 2089
Tamura, N., Murray, G.J., Sharples, R.M., Robertson, D.J., Allington-Smith, J.R.
(2005), OExpr, 13, 4125 (astro-ph/0509913)
Tanaka, M., Yamashita, T., Sato, S., Okuda, H. (1985), PASP, 97, 1020
Thatte, N.A., Weitzel, L., Cameron, M. ET AL. (1994), Proc. SPIE, 2224, 279
Thatte, N., Anders, S., Eisenhauer, F. ET AL. (2000), ASPCS, 195, 206
Thorne, A.P. (1988), Spectrophysics, Chapman & Hall (London)
Tull, R.G., MacQueen, P.J., Sneden, C., Lambert, D.L. (1995), PASP, 107, 251
Vaughan, A.H. (1967), ARA&A, 5, 139
Verheijen, M.A.W., Bershady, M.A., Andersen, D.R. ET AL. (2004), AN, 325, 151
Viton, M., Milliard, B. (2003), PASP, 115, 243
Williams, T.B., Nordsieck, K.H., Reynolds, R.J., Burgh, E.B. (2002), ASPC, 282,
441
Wright, G.S., Rieke, G.H., Colina, L. ET AL. (2004), SPIE, 5487, 653
Wynne, C.G. (1991), MNRAS, 250, 796
Wynne, C.G., Worswick, S.P. (1989), MNRAS, 237, 239
Yee, H.K.C., Ellingson, E., Carlberg, R.G. (1996), ApJS, 102, 269
4. Analysis of 3D data
PIERRE FERRUIT
4.1 Introduction
4.1.1 Presentation and scope
This chapter contains the proceedings of the course on analysis of 3D spectrographic
data given as part of the XVII Canary Island Winter School of Astrophysics. It provides
an overview of some basic and generic analysis techniques for 3D spectrographic data.
It includes a description of an arbitrary selection of tasks with, whenever possible,
examples on real data and a lot of discussion about noise and errors. To illustrate the
examples, we will make a heavy use of tools that are part of the XOasis software developed
at the Centre de Recherche Astrophysique de Lyon (CRAL) and of 3D datasets obtained
using the TIGER and OASIS instruments.
This course is not limited to pure 3D analysis techniques as the core of the analysis of
3D datasets is either identical or similar to what is done for regular spectrographic data.
! It is not exhaustive but contains a rather arbitrary selection of tasks and tools that
It has some obvious caveats and limitations:
! The methods and examples are biased toward extragalactic astronomy.

we have considered as unavoidable.
! It is limited to the data analysis techniques used in visible and near-infrared (NIR)
! It does not address those used in the radio and X-ray communities (long-time users
astronomy.
of 3D spectrography).
4.1.2 Data analysis

Before starting, we need to define better what data analysis is and, in particular, where
it starts and stops in the process leading from raw data to ready-to-publish information
(see Figure 4.1). In a general way, data analysis is considered to follow the data reduction
(processing), which is instrument specific. The analysis will therefore be performed on
radiometrically calibrated datacubes with the instrument signature removed. On output,
it will produce physical quantities (e.g. fluxes, velocities, abundances, etc.).
Note that this is in fact a simplified view as the boundaries between data processing
and analysis are not that well defined. Figure 4.2 shows that many data analysis tools
are in fact at the boundary between data processing and analysis. It also shows that
visualization tools play an important role in the analysis.
4.2 Data, noises, biases and artifacts

4.2.1 Where is the signal?
In a dataset, the ‘signal’ (i.e. what the astronomer will be really interested in) is never
alone. It will be accompanied by undesired backgrounds, noises, artifacts, etc. A signif-
icant part of the work to be performed while analysing the data consists in carefully
extracting the signal and evaluating the associated level of noise. The signal to noise
ratio of a dataset is a key element for its analysis as it will indicate how real, trustworthy
are features detected in a datacube.
In this context, much of the emphasis in this document is on the fact that simply
extracting the signal from a dataset is not enough. One always needs to keep track of the
126
Figure 4.1. Flow chart showing where we locate the data analysis steps in the process leading
from raw 3D dataset to published data.
Figure 4.2. Sketch showing a (non-exhaustive) set of tasks, and an unsuccessful attempt to
sort them between data processing and analysis tasks.
128 Pierre Ferruit
associated noises and how they propagate through the various steps of analysis. We will
use very simple examples to illustrate how this can be done and in the following we will
set the stage and give the basic definitions of the quantities that will be used throughout
this document.
4.2.2 Noises
! the instrument, e.g. through its detector. In the case of the visible and near-infrared
We will distinguish between four different sources of noise in a dataset:
detectors on which our examples will be based, typical detector noises are the read-
! undesired backgrounds like sky emission. Their shot noise will remain even if they
out noise and the shot noise from the detector dark current.
! the signal itself. The signal will have its own shot noise.
have been perfectly subtracted.
! the data reduction. Noise is also introduced in the data during its reduction/
processing, due to the finite signal to noise ratio of the calibration exposures.
Statistical quantities: basic definitions

Given a statistical variable x, we will use
√ the following statistical quantities: its average
x, its variance Vx and its ‘sigma’ σx = Vx . We have the following classical relations:
$N
1
x = ⟨x⟩ = × xi (4.1)
N i=1
$ N
1
Vx = ⟨(x − x)2 ⟩ = × (xi − x)2 (4.2)
N i=1
Application to the noise in a dataset

We will divide the data into four statistical variables: the signal itself, a background, the
dark current and the read-out noise from the detector. This decomposition is typical of
what will be obtained in visible or near-infrared observations.
data = signal + background + dark + read (4.3)
To estimate the associated variance, we will make three assumptions: the four statistical
variables are uncorrelated and their individual variances can be added; the first three have
Poissonian statistical distributions so that their variance corresponds to their average
(Vx = x); the read-out noise of the detector has an average of 0 and a sigma σreadout .
Using these properties we have the following relations:
data = signal + background + dark (4.4)
2
Vdata = signal + background + dark + σreadout (4.5)
The signal to noise ratio of the data is defined as:
S signal signal
=√ =. (4.6)
N Vdata 2
signal + background + dark + σreadout
Noise propagation: basic definitions

If the relations obtained in the previous section are useful to estimate the level of noise in
the raw (unprocessed) data, it will be very important to understand how noise propagates,
i.e. what is gained or lost in terms of signal to noise when applying a data processing or
Analysis of 3D data 129
analysis step. If we have a statistical variable x that is a function of two other statistical
variables y and z (x = f (y, z)), we have the following classical relations (that can be
easily generalized to a function of n statistical variables):
) * ) *
∂f ∂f
xi − x = (yi − y) × + (zi − z) × (4.7)
∂y y z ∂z y z
) * ) *
∂f ∂f
x=y× +z× (4.8)
∂y y z ∂z y z
) *2 ) * ) * ) *2
∂f ∂f ∂f ∂f
Vx = V y + 2 × COVy z + Vz (4.9)
∂y y z ∂y y z ∂z y z ∂z y z
with
$ N
1
COVy z = ⟨(y − y) × (z − z)⟩ = × (yi − y) × (zi − z) (4.10)
N i=1
If the statistical variables are uncorrelated, then the covariance will be zero.
Noise propagation: example

We will apply the noise propagation relations provided in the previous section to the
very simple case where we want to average spectrally the data of a given spectrum of a
datacube. This case corresponds typically to what happens when a datacube is averaged
over a given wavelength range to create an image (see Section 4.3.1). From the signal to
noise ratio in the individual spectral pixels, we will compute the resulting signal to noise
ratio for the averaged data.
Assuming that the spectra are in flux per unit of wavelength, we need to take into
account the size of the spectral pixels when computing the average. We therefore perform
the following operation (for each spectrum in the datacube):
$ N $ N
1
F = × Fi × δλi with ∆λ = δλi (4.11)
∆λ i=1 i=1
Using the relations listed in the previous section, we can derive the average of the sta-
tistical variable F as a function of those of the Fi :
$ N $ N
1
F = × F i × δλi with ∆λ = δλi (4.12)
∆λ i=1 i=1
Here we have assumed that the size of the spectral pixels was perfectly known. If, in
addition, we assume that the Fi are uncorrelated (i.e. their covariance is zero), we have:
N
$ ) *2
δλi
V = Vi × (4.13)
i=1
∆λ
) * 1
,
S F × N F i × δλi
= √ = ∆.λ, i=1 / 0 (4.14)
N V N 2
V × δλi
i=1 i ∆λ
We can immediately see that the contribution of each spectral pixel to the final variance
is weighted by the square of the size of the spectral pixel.
130 Pierre Ferruit
If we now assume that we have a constant spectral pixel size (δλi = cte = δλ; ∆λ =
N × δλ) and that we are limited by the photon noise of the object (Vi = F i ), this yields
the following simplified relations:
1 $
N N
1 $ S √ '
F = F i, V = 2 F i and = N× F (4.15)
N i=1 N i=1 N
We find the classical Poissonian statistics gain in signal to noise ratio proportional to the
square root of N and this example shows how the signal to noise can be tracked when
performing an operation on the data.
Noise propagation: comments

The example of noise propagation given in the previous section aimed primarily at show-
ing how to track noise during a data-processing or analysis step. It also shows that
things can easily get complicated (we had to make a number of simplifying assumptions
to obtain a simple relation). We also point out to the reader that an additional (signifi-
cant) source of complexity is avoided as, in most cases, we only keep track of the variance
of the noise and not of the covariance terms. In particular, resampling operations usu-
ally introduce correlations between data. As a consequence, averaging performed further
down the analysis may not yield the expected increase in signal to noise ratio due to
the covariance terms. It is worth noticing that some data formats allow the storage of
variance or sigma of a dataset. This is the case of the Euro3D data format (Kissler-Patig
et al., 2003), which carries the noise level information for each spectral pixel of each
spectrum in the datacube. We are not aware of 3D data formats providing the user with
a built-in way of keeping track of the complete covariance matrix (of size n2 where n is
the number of spectral pixels in the datacube).
4.2.3 Biases or systematic errors

If noise is unavoidable, at least its importance relative to the signal will decrease most
of the time when a large number of independent measurements are put together. This is
not the case when an error affects all the measurements in the same way. We quite often
refer to this type of errors as systematic errors. One typical example is the error on the
spectrophotometric calibration of an instrument. As the same calibration is applied to
all measurements, they are no longer independent, and they are all impacted in the same
way. This introduces a systematic error.
Biases are another type of error sources that will affect a complete set of measurements.
They are often present when trying to derive properties of a class of objects through
the observation of a sample. In this case, the way the sample is selected is extremely
important and limitations on which object can be observed may introduce biases in the
results. These latter error sources, which can impact the results in a very indirect way,
are usually very difficult to detect and eliminate.
Systematic error: example

In this section, we give a simple example of systematic error and how it can be avoided.
We are placing ourselves in the case of a deep observation of an object made of a sequence
of N exposures of the same field of view. First, we assume that the object is always located
at the same position within the field of view of our instrument, i.e. it falls on the same
spaxels, and the corresponding spectra fall on the same detector pixels each time. For
the sake of simplicity, we will assume that our object corresponds to a single spectrum
per exposure.
During the processing of the dataset, the object spectrum of each exposure is spec-
trophotometrically calibrated by applying a wavelength-dependent factor (which has a
finite accuracy). As the object is always falling at the same location within the instrument
field of view, this factor is the same for all exposures. Once the N spectra are calibrated,
they are then averaged to increase the signal to noise.
To simplify the computation, we will assume that all the exposures are independent and
that the variance of the measurements is the same for all N exposures (i.e. Vi (λ) = V0 (λ)).
We have:
$ N $ N
1 1
F (λ) = × Fi (λ) = × γ(λ) × Γi (λ) (4.16)
N i=1
N i=1
where F (λ) is the final averaged spectrum and γ(λ) is the spectrophotometric correction
factor that is applied to all the uncalibrated spectra of each exposure Γi (λ).
The relation giving the variance of F (λ) as a function of the variance of γ(λ) and of
the variance of the uncalibrated data is as follows:
N ) *2 (N +2
$ γ(λ) Vγ (λ) $
V (λ) = Vi (λ) × + × Γi (λ) (4.17)
i=1
N N2 i=1
Using our simplifying assumptions, this yields:

1 Vγ (λ) / 02
V (λ) = × (γ(λ))2 × V0 (λ) + 2 × F (λ) (4.18)
1N 23 4 1(γ(λ)) 23 4
term decreasing as 1/N constant term
This shows that the spectrophotometric calibration of the data has introduced a system-
atic error term that does not decrease when the datasets are averaged. This means that
after a certain limit, increasing the number of exposures N will not help to increase the
signal to noise ratio. This latter becomes limited by the constant term in the variance.
This constant, systematic error term is due to the fact that the same spectrophoto-
metric correction factor has been used for the spectra of all exposures, introducing a
correlated source of noise in the data. A common way to solve this problem is to place
the object at different locations in each exposure. In this case, the correction factor is
now exposure dependent and becomes γi (λ). We will however assume that its variance
is uniform over the instrument field of view, i.e. Vγ i (λ) = Vγ (λ). The initial relation now
becomes:
$ N $ N
1 1
F (λ) = × Fi (λ) = × γi (λ) × Γi (λ) (4.19)
N i=1
N i=1
And this yields:

$N N
1 2 Vγ (λ) $ / 02
V (λ) = 2
× V i (λ) (γi (λ)) + 2
× Γi (λ) (4.20)
N i=1
N i=1
1 23 4 1 23 4
term decreasing as 1/N term decreasing as 1/N
The constant term in the variance has now disappeared and has been replaced by a term
decreasing as 1/N . By changing the observation strategy, it has been possible to avoid
introducing a systematic error term. This illustrates how keeping track of the noise can
help in devising better observation, processing or analysis strategies.
132 Pierre Ferruit
Figure 4.3. Sketch illustrating how an artifact can be spread over neighbouring spaxels when
resampling spatially a dataset.
4.2.4 Artifacts
Artifacts are one last item in our list of features that will affect data. Typical sources of
! the detector (bad columns, bad pixels, etc.);

artifacts in visible and near-infrared datasets are:
! the cosmic ray impacts on the detector and the residuals of their removal; and
! the residuals from the sky subtraction (see Section 4.4).
As for the noise, it is important to keep track of pixels affected by artifacts. In the Euro3D
format, this is done through the use of a data quality flag for each spectral pixel of each
spectrum of the datacube (see Section 4 of Kissler-Patig et al., 2003).
The artifacts and the resampling of data

When resampling the data (spectrally and/or spatially), artifacts can be spread over sev-
eral spaxels and/or spectral pixels. This is one of the main reasons why resampling of data
is quite often seen as evil by many astronomers (another reason being that it introduces
correlation between one pixel/spaxel and another). This is illustrated in Figure 4.3. Unfor-
tunately, resampling of data in wavelength is usually quite unavoidable during the data
processing (wavelength calibration). This is not the case of spatial resampling that can
quite often be avoided (and usually is) until the very last stages of the data processing/
analysis.
4.3 Image reconstruction and datacube visualization

4.3.1 From a cube to an image
Displaying images and spectra extracted from a datacube remains one of the most efficient
ways (or at least one of the most used methods) of exploring a datacube. However, the
reconstruction of an image by collapsing a datacube along the wavelength axis is not
completely straightforward. The main reason for that is the fact that many datacubes
do not have a regular spatial sampling.
! Many 3D spectrographs have sampling patterns that are not a regular grid (e.g. in
The spatial sampling provided by the spectrograph can be irregular in different ways:
! The sampling of the field of view can be sparse.

the case of circular or hexagonal spaxels).
! There can be holes in the field of view coverage even when the sampling is not
intrinsically sparse (observation that is the sum of non-contiguous fields of view,
rejected spaxels, etc.).
Figure 4.4. Images showing the location of the spaxels within the field of view of three different
3D spectrographs (SAURON, OASIS and PMAS in its PPAK mode).
Figure 4.5. Example from the image reconstruction and visualization of a datacube with-
out (left panel) and with (right panel) spatial resampling. The figure has been extracted from
Sánchez (2004) and corresponds to a dataset of the INTEGRAL 3D spectrograph, which has
circular, non-contiguous spaxels.
Examples of the spatial distribution of spaxels for various 3D spectrographs are shown
in Figure 4.4.
When reconstructing an image from a datacube, two different approaches can be used
depending on the adopted philosophy about the resampling of data. To avoid resampling,
a special visualization tool like the E3D software (see tutorial on the use of E3D in this
book) can be used that will take into account the location and shape of the spaxels and
display them without resampling (see left panel of Figure 4.5).
134 Pierre Ferruit
Figure 4.6. Example of image reconstruction procedure using the XOasis software.
The right panel of Figure 4.5 shows the results obtained for the same dataset when
resampling and interpolation of the data are used (the black circles superimposed on the
image are the shape and position of the spaxels). In Figure 4.6, we show screenshots of
the graphical user interface (GUI) associated to the image reconstruction tools of the
XOasis software suite. It clearly shows the steps to be followed from a datacube to an
! The first step is the ‘collapsing’ of the spectra along the wavelength direction (upper
image:
panel of Figure 4.6). The user defines a wavelength range (that can be the whole
spectrum) over which the spectral pixels will be averaged or summed (possibly using
weighted summation). As an output, a table is produced that – for each spaxel of
the datacube – contains its position and the value obtained when collapsing the
! The second step (central panel of Figure 4.6) is the actual image reconstruction
corresponding spectrum.
that will produce an image with square pixels and a regular sampling from the table
produced during the first step. A triangulation will first be performed (looking for the
nearest neighbours of each new sampling point in the table) and will be followed by
the interpolation itself. The user is provided with the choice of allowing extrapolation
of data and/or filling in the large gaps between spaxels by interpolation. In the
example we have chosen extrapolation was not allowed, and the output image values
outside the original field of view are filled with zero values (see bottom panel of
Figure 4.6).
In much software (e.g. E3D) the user is also provided with a choice of interpolation
methods.
4.3.2 Datacube visualization

The question of datacube visualization itself will only be briefly assessed in this document
as it is the subject of a dedicated tutorial in this book. The art is, of course, to be
able to display a 3D datacube on a 2D display. As mentioned at the beginning of this
chapter, the simplest and most common solution is to reduce the number of dimensions
of the data to be displayed by either collapsing the datacube along the wavelength axis
(image representation; from 3D to 2D), or extracting individual spectral by collapsing
the datacube along the spatial directions (spectrum representation; 3D to 2D or 1D).
The E3D software is a very good example of how to navigate between spectrum and
image representations of a datacube.
In the radio community, the side-to-side display of velocity slices (channel maps cor-
responding to an image representation) is commonly used. It is also possible to display
position–velocity diagrams in which only one spatial axis of the data is preserved and
displayed together with the wavelength information. This representation is familiar to
the users of long-slit spectrographs.
However, none of these visualization schemes is actually a true 3D visualization
method. An example of a ‘real’ 3D visualization developed and used by J. Gerssen is given
in Figure 4.7. It has been used for the visualization of a Lyα datacube (see Wilman et al.,
2005). This type of visualization is best used when displaying well-defined features like
emission lines. It is usually difficult to extract quantitative information from these views.
4.4 Getting rid of atmospheric effects and undesired backgrounds

Remaining at the borderline between data processing and analysis, we will now focus on
the removal of atmospheric effects and undesired backgrounds.
4.4.1 Sky emission, zodiacal light and thermal background

If we do not consider instrument-related backgrounds like stray light or thermal emission
from the optics in the infrared (IR), the most common (and unavoidable) background
source for ground-based observations is the atmosphere.
136 Pierre Ferruit
Figure 4.7. Snapshot of datacube visualization using the VolView tools extracted from the
work of J. Gerssen.
20
18
16
Flux density (phot/s/nm/arcsec 2/m 2)
14
12
10
0 0 1 2 3 4 5 6
350 450 550 650 750 850 950 1050 1150
Wavelength (nm)
Figure 4.8. Spectra of the sky emission in the visible (left panel, from the GEMINI observatory)
and of the zodiacal light (right panel).
Obviously, all the atmospheric background will disappear in the case of observations
in space. The dominant external background then becomes the (much fainter) zodiacal
light originating from the dust localized in the ecliptic plane of the Solar System. It is
a mix of scattering of solar light (toward the blue) and thermal emission from the dust
(toward the red). Typical sky and zodiacal light spectra are shown in Figure 4.8.
4.4.2 Subtracting a background

In order to get the signal out of the data, it will be necessary to subtract these back-
grounds. However, even a perfect background subtraction will not remove the associated
noise. In the following, we give a very simple example to illustrate the impact of the
background noise on the final signal to noise ratio.
We will assume that we have data T obtained at the location of an object as well as
data Ti obtained specifically for background subtraction at a set of N different locations.
We will have:
T = S + B + d and Ti = Bi + di (4.21)
where S is the signal, B and Bi the background and d and di the detector dark current.
The associated variances are (assuming a ‘shot-noise’ type of contributions for the signal,
background and dark current):
i i
VT = S + B + d + σ 2 and VTi = B + d + σ 2 (4.22)
where σ is the detector read-out noise term.

The first step toward isolating the signal S is to subtract the dark current (that could be
considered as an instrumental background). This will be performed using a dark current
map and we will assume that the relative accuracy of this map is uniform (δd/d =
i i
δdi /d = cte), as is the dark current rate (d = d ). This yields the following relation for
the variance associated to the dark current subtraction:
) *2
δd
Vd = V d i = d × (4.23)
d
And we have:
) *
δd2
(T − d) = S + B and VT −d = S + B + 1 + 2 × d + σ 2 (4.24)
d
) *
i i δd2
(T − d) = B and VT −d = S + B + 1 + 2 × d + σ 2
i i i
(4.25)
d
The second step is the background subtraction. For this purpose, we will average all
the background-only measurements B i = (T − d)i and subtract the result to S + B =
i
(T − d). Assuming that the background is uniform over the field of view (B = B = cte),
we have:
$N $N
1 1
S = (T − d) −
× B i = (T − d) − × (T − d)i (4.26)
N i=1
N i=1
) * ) ) 2
* *
1 δd
V =S+ 1+ × B + 1 + 2 × d + σ2 (4.27)
N d
The multiplicative factor 1 + 1/N witnesses the increase in variance associated with the
background subtraction. For low N values, this latter can have a significant impact on
the signal to noise ratio of an observation.
4.4.3 Where things can get even more complicated

Beware of instrumental effects
Obtaining the background and object spectra using different optical paths, as is often
! The instrument spectral point-spread function (PSF) can differ from one spaxel to
the case, can be a problem. This is mainly due to two different effects:
the other. As an example, the spectral PSF of an integral field unit (IFU) quite often
! The wavelength calibration of a datacube is not perfect and there may be shifts
varies within its field of view.
between spaxel spectra.

138 Pierre Ferruit
Image:
Image: template
templete Image: subwide Image: subshiftred
0.4 0.004
0.002
Pixel value (none given)

0.3 0.001
0.002
0.2 0
0 −0.001
0.1
−0.002
0 −0.002
−4 −2 0 2 4 −4 −2 0 2 4 −4 −2 0 2 4
Position ( ) Position ( ) Position ( )
Figure 4.9. Left panel: plot of the nominal Gaussian line profile with the plots of the narrow/
broader and blueshifted/redshifted profiles superimposed. Central panel: residuals from the sub-
traction of the narrower/broader profiles from the nominal one. Right panel: residual from the
subtraction of the blueshifted/redshifted profiles from the nominal one.
Figure 4.10. Left panel: spectrum of an active galaxy in the 630–700 nm range displaying the
prominent emission lines from the object and an atmospheric absorption feature between 680
and 690 nm. Right panel: zoom on the atmospheric absorption feature (which is made of several
lines).
In the case of the subtraction of a background with a continuum spectrum, these effects
will have little impact. They can however be a source of artifacts when subtracting a
background with spectrally unresolved emission lines (i.e. the sky).
To illustrate this effect, we have performed a simple simulation. We have produced
an artificial emission line with a Gaussian profile (centred on zero and with a sigma
of unity; see Figure 4.9). We have then introduced very small changes in the Gaussian
profile parameters (changes in sigma of 1%; shifts of the centroid of the line by 1% of
the sigma of the Gaussian profile) and created the corresponding spectra. A plot with all
three profiles overlapped is displayed in Figure 4.9, and it shows that the lines are barely
distinguishable.
However, if we now subtract the lines with the modified Gaussian profile to the original
one, we see that we end up with negative and positive residuals that can reach 0.5–1.0%
of the peak intensity of the line (see Figure 4.9). Although the amplitude of the residuals
seems small, this can be a major issue when the background is stronger than the object,
as it is quite often the case for deep exposures and in the near-infrared.
The case of atmospheric absorption features

So far, we have only addressed the subtraction of additive backgrounds. In the case of
ground-based observations, the sky background can indeed be considered as an additive
background, but it is also associated with absorption features (see Figure 4.10). They
can be a nuisance and are quite difficult to correct for.
Figure 4.11. Sketch illustrating how the spectrum of a point source can be distorted due to
the wavelength-dependent shift induced by atmospheric refraction.
One common way to determine the correction for this absorption is to calibrate it
using a featureless spectrum such as the spectrum of the spectrophotometric standard
star used for the radiometric calibration of the instrument (this latter being quite often
! Assume the stellar continuum is flat at the location of the absorption feature.
selected to have a spectrum as featureless as possible). The steps are as follows:
! Fit the absorption line and determine the relative absorption (with respect to the
! Apply the corresponding multiplicative correction to all the spectra of the datacube.
continuum level).
A major caveat is the fact that the importance of this absorption feature depends strongly
on the parameters of the observation (in particular the time and the zenithal angle), and
they can be quite different for the spectrophotometric star and the object itself. It is
therefore not usually possible to apply the correction directly, and one needs to scale it
to the object airmass and hope that time variation of the absorption between the two
observations has remained small.
Atmospheric refraction
One last atmospheric effect is atmospheric refraction. It will cause the image of an object
to be shifted as a function of wavelength. The effect decreases as wavelength increases
and it is usually negligible starting in the near-infrared.
In slit spectroscopy it will cause the object to gradually move outside of the slit as
wavelength changes. In 3D spectroscopy, there is no longer a risk of ‘losing’ the object (as
long as the field of view is large enough) as it will just ‘move’ within the field of view as a
function of wavelength. Reconstructing monochromatic images at different wavelengths
and comparing them will easily show the shift. However, if it is not corrected, this shift
will distort the spectra of the object, changing its slope from one spaxel to the other, as
illustrated in Figure 4.11.
A detailed description of the impact of atmospheric refraction and how to correct for
it can be found in Arribas et al. (1999). Typically, the steps of the correction procedure
are the following (and they are only possible when using a 3D spectrograph):
140 Pierre Ferruit
! Divide the datacube in monochromatic ‘slices’ (one image of the field of view per
! Shift all slices by the correct amount.

spectral pixel).
! Resample all the slices spatially using a common sampling grid (see comments on
! Reconstruct a datacube by stacking the corrected slices.

resampling data in Section 4.2.4).
One difficulty can be the determination of the correct shift. It is possible to use theoretical
values or to use a point source within the field of view (if available), paying great attention
not to measure shifts intrinsic to the source itself.
Note that some instruments are now equipped with atmospheric refraction correctors
(ADC). This is almost mandatory when using 3D + adaptive optics in the visible.
4.5 From a collection of datacubes to a single one

A complete observation quite often consists in a collection of datacubes and in this
chapter we will discuss the different steps that will lead from this collection to a single
datacube. We will distinguish between different observing strategies that have different
! Obtaining a collection of observations of the same field of view with only a small
goals:
pointing offset between each exposure in order to improve the spatial sampling of
! Obtaining a collection of observations of the same field of view to go as deep as

the datacube, i.e. the so-called ‘drizzling’ techniques.
! Obtaining observations of a collection of neighbouring (although sometimes non-

possible (to improve the final signal to noise ratio).
contiguous or overlapping) sub-fields in order to increase the observed area on the

sky (mosaicking technique).
Note that we will not address the question of datacubes obtained on the same field of
view but within different spectral ranges.
4.5.1 Drizzling
The drizzling technique has been advertised a lot within the astronomical community
through its use for HST (Hubble Space Telescope) images. It aims at improving the
sampling of the image and can be very useful for a diffraction-limited instrument oper-
ating over a large wavelength range. In this case, the size of the PSF will change quickly
with wavelength, and quite often the instruments cannot sample it correctly at all wave-
lengths and the PSF ends up undersampled in the blue; see Hook and Fruchter (2000)
for a description of the technique and its application to images.
So far it has only been applied to images but it can be naturally extended to 3D
datacubes (after all, they can be considered as stacks of monochromatic slices). However,
we are not aware of any drizzling software adapted to datacubes.
4.5.2 The deep-field approach

In the deep-field approach, multiple exposures of a single field of view are obtained,
aiming at pushing the detection limit of an instrument. Systematic dithering of the
exposures is usually performed in order to facilitate the elimination of artifacts and
limit the introduction of systematic errors. The observation strategy for datacubes is, in
essence, similar to the one used for the imaging deep-fields.
The basic minimum steps needed to co-add (merge) the deep-field datacubes are: get
the datacubes; register them with respect to each other, normalize them with respect to
each other; co-add (merge) them.
Figure 4.12. Reconstructed images for the 7 individual NGC 4258 datacubes (to the left)
obtained with the OASIS/CFHT instrument and for the final datacube (to the right) obtained
by co-adding (merging) them.
We will discuss some of these points in the following subsections and we will systemat-
ically use some OASIS/CFHT (Optically Adaptive System for Imaging Spectroscopy /
Canada–France–Hawaii Telescope) observations of the galaxy NGC 4258 as an example
(see Figure 4.12).
Relative positioning
Before trying to co-add the datacubes, it is necessary to determine their relative posi-
tioning. The easy way is of course to know the offset beforehand. This is rarely possible,
especially as many observations have sub-arcsecond samplings and it is necessary to reach
accuracies of a fraction of a spaxel size. Quite often, this information must therefore be
obtained from the data itself.
If a ‘sharp’ morphological feature (e.g. the nucleus of a galaxy, a star, etc.) is available
within the field of view it is possible to use the position of this source (as derived,
for example, from centroiding or Gaussian profile fitting) as a reference. If no ‘sharp’
morphological feature is available, it is also possible to use contour plots of reconstructed
images to get the relative positioning of two datacubes. This is illustrated in Figure 4.13.
! As it is, with inspection of the overlap being done by eye, this is a very subjective
The difficulties are usually as follows:
! Changes in observing conditions (e.g. a seeing change) will mess up with the contour
method.
! Noise in the individual exposures does not help.

superposition.
Relative normalization
After registering the datacubes with respect to each other, we also have to check their
relative normalization, i.e. tracking differences in radiometric response of the complete
system (including the atmosphere if relevant) from one observation to another. In most
cases, a dedicated spectrophotometric calibration will not be available for each exposure
142 Pierre Ferruit
Frame: Halphala Ident: Frame: Halphala Ident:

6 6
4 4
2 2
Position
Position
0 0
−2 −2
−4 −4
−6 −6
−5 0 5 −5 0 5
Position Position
Figure 4.13. Example of the use of contour plots of a reconstructed image to derive the relative
positioning of two datacubes. Left panel: superimposed contours of the two images for the initial
positioning. Right panel: same but for the final positioning.
and it is very difficult to track changes like a modification of the transparency of the
atmosphere. In the case of the deep-field approach, getting the relative normalization
between the datacubes is straightforward as they all cover the same field of view and
their flux or surface brightness levels can easily be compared.
Ready to co-add?
At this stage, we should have N datacubes that have been registered and normalized
with respect to each other and we should be ready to co-add (merge) them. However,
! How should we co-add them?

before starting the process we need to answer the following questions:
! Is it optimal in terms of signal to noise (which is the driver for the deep field
approach)?
In order to answer these questions, we are going to use a very simple and quite arti-
ficial example. It will show that performing a weighted summation when co-adding the
datacubes can help a lot. This allows us to take into account the fact that the datacubes
can show up differences in signal to noise ratio and reflects the quite obvious fact that it
is not desirable to give the same weight to low and high signal to noise datacubes.
We will assume that the calibration of the datacubes has introduced a division by the
integration times (to get the spectra in unit of power, i.e. energy per unit of time) and
that we are in photon noise regime. This yields:
) *i √
1 i S
i
Scalibrated = i × S i and V i = S and = Si (4.28)
t N
The variance of a calibrated datacube is:
) *2 ) *2
i
∂Vcalibrated i 1 1 i
i
Vcalibrated =V ×
i
=S × i = i × S calibrated (4.29)
∂V i t t
If we co-add (merge) the datacubes using a straight-forward average (no weights), we

have:
$ N
1
Scalibrated = × i
Scalibrated (4.30)
N i=1
If we assume that all the calibrated signals have the same average value S calibrated =
i
S calibrated (further simplification), this yields the following relations:
$ N N
1 S calibrated $ 1
Vcalibrated = 2
× i
Vcalibrated = × (4.31)
N i=1
N2 i=1
ti
) * .
S S calibrated 1
=√ = ., ×N × S calibrated (4.32)
N calibrated Vcalibrated N 1
i=1 t i
1 23 4
source of trouble
The sum of the inverse of the integration times that appear in the equation for the signal
to noise ratio is going to be a source of trouble. To see that, we are going to compare
two practical cases.
First, if all the integration times are the same (ti = t = cte):
) * . .
S N √ √
= . × S calibrated = t × N × S calibrated (4.33)
N calibrated N 1 23 4
t usual factor
Second, if we have two exposures with different integration times (ti=1 = t and ti=2 =
t/2):
) * . .
S 2 √ 2
= . × S calibrated = t × √ × S calibrated (4.34)
N calibrated 3 3
t 1 23 4
smaller factor
We can see that for exposures of different integration times, we do not have the expected
gain in signal to noise ratio (expected to be proportional to the square root of the
cumulated integration time). This is due to the contribution of the short exposure time
exposures. The solution is to put more weight on long exposures than on short ones.
With ttot being the sum of all individual integration times (cumulated integration time)
and using a normalized weight ti /ttot , we have:
N
$ N
$
weighted 1 1
Scalibrated = × i
t × i
Scalibrated = × Si (4.35)
ttot i=1
ttot i=1
N
$ N
$
weighted 1 1
Vcalibrated = × Vi = × Si (4.36)
t2tot i=1
t2tot i=1
) *weighted weighted .
S S √
= . calibrated = ttot × S calibrated (4.37)
N weighted
calibrated Vcalibrated
We have obtained the classical factor of square root of the cumulated integration time.
This simple and naive example outlines the fact that it is necessary to assess the impact
of each processing/analysis step on the signal to noise of the data. In this case the simple
fact of having calibrated data has forced us to modify the way the datacubes are co-added
(if data had still been in counts, a direct average would have worked). This explains as
an example why the HST images are never in physical units (per unit of time) by default.
It is simply to allow the user to perform direct summations on them without having to
worry about weights.
144 Pierre Ferruit
Figure 4.14. Example of graphical user interface (from the XOasis software) used as a front-
end for a datacube merging tool. It can be seen that this program allows use of a weighted
summation (with the weights set by default to ti /ttot ).
Figure 4.15. Spectra extracted from an individual datacube (left panel) and from the final
merged datacube (right panel) at the same location in the object.
The (very simple) procedure we have described is in fact quite similar (although less
sophisticated) than the ones used in optimal summation techniques.
A simple illustrated example

To finish on an example, we show a snapshot of the graphical interface of the ‘merge
datacube’ task of the XOasis software in Figure 4.14. Note that in this case we have simul-
taneously corrected for atmospheric refraction. Coupling this correction with the merging
of the datacubes allows for a limiting of the number of spatial resamplings of the data.
In Figure 4.15, we also display a comparison between a spectrum extracted from an indi-
vidual datacube and a spectrum extracted from the final, merged datacube. The gain in
signal to noise ratio is obvious.
Back to the deep-field approach and having a look at what is ahead of us
In the previous sections, we have used simple and fairly artificial examples to illus-
trate how to co-add the exposures for a deep-field type of observation. A real example
of 3D spectrographic deep-field can be found in Wilman et al. (2005). It has been per-
formed using the SAURON (Spectrographic Areal Unit for Research on Optical Nebulae)
instrument, with a cumulative integration time of more than 15 hours (combination of
! having a very good and deep knowledge of the instrument (well-calibrated

30 individual exposures). The main difficulties that have been identified are:
! eliminating the backgrounds and systematic errors;

and -characterized instrument);
! coping with changes of image quality and atmospheric transmission from one expo-
sure to another. Note that how to deal with changes in image quality is an entire
! coping with the intrinsic stability of the telescope plus instrument system. This is
subject in itself and is not addressed here;
critical for this type of observation, which tends to push the instruments to their
limits.
This type of prototype study is paving the way for future 3D spectrographic deep fields
that will be at the heart of the scientific goals of the future generation of 3D spectrographs
like MUSE (Multi-Unit Spectroscopic Explorer) and KMOS (K-Band Multi-Object Spec-
trometer) for the VLT (very large telescope). Dedicated software will be needed for
the co-addition of these tens of cubes (cumulative integration times that can reach
80 hours).
4.5.3 Mosaicking
In the mosaicking approach, a large number of (overlapping or not) sub-fields are obtained
and are then merged to create the final field of view. In the following, we will outline the
differences with the deep-field approach that has been described in detail in the previous
section and we will give examples of mosaics of 3D spectrographic fields.
Differences with the deep-field approach

The observing procedure to obtain a mosaic of field is very similar to the one used for
! one or two observations are obtained per field of view. If two are obtained, a small
the deep-field approach, but for the offsets between the fields:
offset will be applied between them in order to facilitate the elimination of artifacts
! known offsets are applied between fields of view. Depending on the scientific goals
and to limit the introduction of systematic errors;
and on the observation constraints, these individual fields can be contiguous or not,
and the degree of overlap can vary.
The relatively small number of exposures per field of view compared to the case of
the deep field makes the problem of systematic errors much less critical. However, the
registration (see Section 4.5.2, Relative positioning) and normalization (see Section 4.5.2,
Relative normalization) steps are more troublesome in the case of a mosaic of fields.
Relative registration and normalization of the exposures

The problem comes from the lack of overlap between the fields of view. The methods for
the relative registration and normalization of the exposures described for the deep-field
approach rely heavily on the comparison of data obtained on the same patch of the sky.
In the case of a mosaic of fields like the one shown in Figure 4.16, it is clear that:
146 Pierre Ferruit
The barred galaxy NGC 2903

0 10−19 W m−2 Å−1 arcsec−2 10 0 10−19 W m−2 arcsec−2 350 −200 km s−1 +200
Martin et al.
60 60 60
3D spectrography
40 40 40
[NII]+Ha+[SII]
20 20 20 emission-lines
+
Arcsec
0 0 0
continuum
−20 −20 −20
−40 −40 −40 mosaic of

20 fields
−60 Continuum −60 Hα −60 Velocity
(15” × 15” each)
−80 −80 −80
a total of
−40 −20 0 20 40 −40 −20 0 20 40 −40 −20 0 20 40
Arcsec Arcsec Arcsec 19908 spectra
OASIS/CFHT - February 2000
Figure 4.16. Reconstructed images of NGC 2903 for a mosaic of 20 sub-fields obtained with
the OASIS/CFHT instrument.
! The number of spatial feature that could be used for registration purposes is very
! Any normalization would have to be propagated from one datacube to the next,
small and, quite often, they appear in a single sub-field (no overlap).
quite often using very small regions of overlap and yielding large errors.
A common solution is usually to make sure that a wide-field image in the same wavelength
range is available. It is then possible to systematically compare reconstructed images to
this wide-field image and derive the necessary registration and normalization information.
This allows for a single common reference for all the sub-fields but does not solve the
problems of the lack of reference sources for the registration in some sub-fields and the
problems of changes of observing conditions from one sub-field to another (heterogeneous
dataset).
Simple illustrated examples

In Figure 4.16 and Figure 4.17, we give two simple examples of mosaic of sub-fields
observed in 3D spectrography. The first one on NGC 2903 is a prototypical example of
a mosaic of a large number of non-overlapping sub-fields, yielding a sparse, incomplete
coverage of a very extended field of view. As described in the previous section, a wide-
field image has been used to register and normalize all the individual datacubes. The
final, merged datacube contains more than 19,000 spectra.
The second example on NGC 1068 is intermediate between a mosaic and a deep-field.
Offsets have been applied between the sub-fields, but a large overlap has usually been
preserved to increase the cumulated integration time on some regions. This type of obser-
vation is extremely sensitive to changes in seeing conditions. Due to the complex structure
of the object, these changes will introduce modifications in the observed morphology that
cannot be corrected by a simple normalization. Prominent artifact – usually located at
the boundaries of the sub-fields – can then appear when co-adding (merging) this set of
heterogeneous datacubes.
Last, we outline in the right panel of Figure 4.17 that additional artifacts can be
introduced by the merging software. It is important that this software does not interpolate
between sub-fields if the distance between the points is too large. In this example, the
Figure 4.17. Example of a mosaic of fields obtained using the OASIS/CFHT 3D spectrograph
on the active galaxy NGC 1068.
‘no-fill’ option of the XOasis ‘merge datacubes’ tool has been disabled, leading to the
creation of non-physical interpolated spaxels in the final datacube (and they are not that
easy to spot due to the quality of the interpolation).
4.6 Smoothing and binning

4.6.1 Smoothing data
Smoothing of data consists in convolving the data with a spatial or spectral filter. As for
the interpolation, numerous different methods exist. The most straight-forward one is the
‘boxcar’ type of smoothing, where a pixel/spaxel is averaged with its closest neighbours.
Most of the time smoothing is applied in order to increase the signal to noise in the data.
In terms of noise coupling, it has effects equivalent to those of resampling data (pixels or
spaxels are not independent anymore and their noises are correlated).
We will not develop the topic of smoothing further as it is usually better to use a
good binning technique (see next section) instead. In fact, the binning method presented
below can be seen as an advanced smoothing technique.
4.6.2 Binning
Binning has the same objectives as smoothing and consists of putting together different
pixels/spaxels in order to increase the signal to noise in the data. The main difference
with smoothing is that it actually reduces the number of pixels/spaxels (there is no data
duplication). In the following, we will concentrate on spatial binning.
148 Pierre Ferruit
4
4050
2
2
V (km/s)
Arcsec
Arcsec
4000
0
0
−2
−2
3850
−4
−4
−4 −2 0 2 4 −4 −2 0 2 4 −4 −2 0 2 4
Arcsec Arcsec R (Arcsec)
3920 km/s 4000 km/s 4070 km/s 180 km/s 200 km/s 260 km/s
Figure 4.18. Example of kinematical maps reconstructed from a datacube binned using the
technique developed by Cappellari and Copin (2003) (courtesy of G. Dumas).
The main difficulty is the fact that the signal to noise ratio within a datacube is not
uniform. As a consequence, if binning is useful in some regions of a field of view (e.g. the
external regions of a galaxy) it may not be needed in others (e.g. the core of the galaxy
that would be present in the same observation). And, as spatial binning will decrease the
spatial resolution, binning may not only be useless in some regions but also unwelcome.
This has led to the emergence of the so-called adaptive binning techniques like the
one based on Voronoi tessellation developed by Cappellari and Copin (2003). These
techniques are called adaptive since the size of the bins is tuned in order to ensure the
signal to noise ratio in the binned datacube is as uniform as possible. This means that
spaxels with a high enough signal to noise ratio will remain untouched while the binning
will only be applied to areas where it is needed. The main constraints on the bins used
! The bin shall not overlap (this constraint is not satisfied by the smoothing techniques)
by Cappellari and Copin (2003) are the following:
! In order to try to avoid exotic bin shapes, the bins shall be as ‘compact’ (round) as
and no holes shall be introduced.
possible.
In their paper, these two authors present a method satisfying these constraints and its
practical implementation. It has been widely used in the SAURON project, and Fig-
ure 4.18 shows an example of maps reconstructed from an adaptively binned datacube.
4.7 Data mining and crowded-field spectrophotometry

4.7.1 Data mining
Data mining could be defined as the art of detecting small and weak objects within a
huge datacube. It is becoming more and more important with the development of large
3D spectrographs and the increasing importance of spectrographic deep fields. Solutions
! by the X-ray community for the automated detection of X-ray sources in surveys
have already been explored in various astronomical communities:
! by the radio community for the detection of sources in the HI surveys as an example.
(with a 3D touch as they can have energy-sensitive detectors).
A quite interesting example of data mining application of HI self absorption features
! for the detection of sources and construction of catalogues in wide-field imaging

in a survey can be found in Gibson et al. (2005).
surveys; see, for example, the SExtractor software from Bertin and Arnouts (1996).
Figure 4.19. Example of application of the crowded field spectro-photometry with sources (the
four quasar images) located on top of a non-uniform background (the lensing galaxy) (courtesy
of M. Roth).
In the optical and infrared 3D community, these techniques are only just starting to
emerge, and in the last few years work has been conducted by some members of the
Euro3D network (e.g. S. Foucaud and G. Covone). Two prototype tools have been devel-
oped: one based on the systematic use of SExtractor on monochromatic slices of a cube,
and the other based on the detection of correlations between spectra.
4.7.2 Crowded-field spectrophotometry

Another expanding domain is the field of crowded-field spectrophotometry, which aims
at extracting the spectrophotometry of a collection of sources within a ‘crowded’ field
of view (e.g. see example in Figure 4.19). A lot of work has been conducted in this field
by the team at AIP (Astrophysikalisches Institut Potsdam), and the applications are
in fact not limited to surveys of numerous sources but can also be efficiently applied
to observations of single objects (planetary nebula, supernova, etc.). As in the case of
! How to efficiently and cleanly get rid of the backgrounds, which are usually not
imaging, the major issues are:
! How to efficiently collect the flux of an object that is spread over several spaxels and
uniform spatially?
may be contaminated by the flux of neighbouring objects?

In particular, for the second point, an accurate knowledge of the PSF is key for an
accurate spectrophotometry.
4.8 Model fitting

After a few chapters meandering at the frontier between processing and analysis, we really
enter the domain of pure data analysis. We also enter the domain where things become
extremely specialized (depending on the type of objects and the scientific goals). The
model can in fact be extremely sophisticated both technically (e.g. when some models
are used to produce synthetic datacubes that are compared to the observed ones) and
150 Pierre Ferruit
scientifically. In the following, we will focus on two simple types of model fittings: the
fitting of the stellar continuum and of emission lines.
4.8.1 The example of stellar continuum fitting

In extragalactic observations, the study of the stellar continuum of the objects allows
extraction of information about their stellar contents (e.g. to derive the stellar population
and also to trace back the history of the stellar formation) and their kinematics. The
techniques used for stellar continuum fitting in 3D spectrography have been directly
adapted from those used in regular spectrography. In this domain, a lot of work has
been conducted by the SAURON team (amongst others), and we will present briefly the
pixel-fitting technique developed by members of the this team.
These developments have been driven by the fact that usual methods like Fourier
Correlation Quotient (FCQ) method (Bender, 1990) work in Fourier space and are very
sensitive to the presence of artifacts in the spectra or simply of emission lines. In fact,
this last point turned out to be an extremely strong limitation, and it led them to develop
further a method, called pixel-fitting, which is closer to a direct fit of the spectra. It is
based on the minimization of the difference between the observed spectra and a template
spectrum shifted and broadened. Pixels with artifacts or regions where emission lines can
be found are simply rejected.
The procedure used to determine both the stellar contents and the stellar kinematics
! Using a single stellar template spectrum, a first guess of the stellar kinematics of the
is iterative and as follows:
! Using this first guess as a constraint (i.e. the kinematics is kept fixed during the fit),
object is derived.
the best linear combination of stellar template spectra is determined (determination
! The first step is then repeated, this time using the optimal template, yielding a new,
of the ‘optimal’ template).
refined estimate of the stellar kinematics.

Further iterations are possible but are usually not necessary.
Establishing which regions should be rejected during the fit due to emission lines is
usually difficult, and they are usually oversized in order to be applicable to the complete
field of view (it would be cumbersome to examine each spectrum individually to adapt
the size of these exclusion windows). Given that, the SAURON group and in particular
J. Falcón-Barroso went slightly further and they have coupled the fitting of the stellar
continuum and the emission lines (single Gaussian profile, mainly the Hβ and [OIII] lines)
in the SAURON data. They have therefore developed the tools to perform the optimal
template and emission-line fitting simultaneously; an example of fit on a single spectrum
is shown in Figure 4.20 and examples of maps derived using this pixel-fitting software
can be seen in Falcón-Barroso et al. (2006).
Although this tool has proven very efficient and useful in the case of the galaxies
observed with SAURON, it starts to reach its limits when the emission lines become
strong compared with the continuum and when they start to display complex profiles
that cannot be described by a single Gaussian profile. In these conditions, it is difficult
to avoid fitting the stellar continuum and the emission lines independently, each with
dedicated software.
4.8.2 A step-by-step example of emission-line fitting

As outlined above, it is sometimes unavoidable to fit the stellar continuum and the
emission lines independently. As for establishing the stellar contents and kinematics, it
is possible to proceed iteratively. The corresponding workflow is shown in Figure 4.21.
Figure 4.20. SAURON spectrum showing the superposition of the stellar continuum and of
emission lines (top panel) and the corresponding optimal template spectrum (linear combination
of stellar templates from a library) as determined using the pixel-fitting software developed by
J. Falcón-Barroso (courtesy of J. Falcón-Barroso).
Figure 4.21. Typical workflow for the fitting of the stellar continuum spectrum and of the
emission lines when they cannot be fitted simultaneously.
152 Pierre Ferruit
Image: specBN Row: #1
20
15
Pixel value
10
0
6300 6400 6500 6600 6700 6800
Position
Figure 4.22. Typical spectrum of the central region of NGC 2110 that will be used to prepare
the fit of the complete datacube. It includes lines from [OI] (doublet), [NII] (doublet), Hα and
[SII] (doublet).
In the following we will describe an example of fitting of emission lines in a continuum

subtracted datacube (i.e. a small part of this workflow), based on the use of the ‘fit/spec’
software developed by A. Pécontal (Rousset, 1992).
This tool was developed for the fitting of 3D observations of active galaxies and has
a philosophy very similar to the tool presented in the tutorial of this Winter School. It
allows the fitting of single or multiple lines using single or multiple Gaussian profiles.
Systems and constraints

In many cases, fitting the emission lines with a single Gaussian profile is not enough and
it is necessary to use multiple ones. This increases the number of free parameters and the
fitting process becomes more complex and also much less stable. If we add to that the
fact that fitting is usually conducted over hundreds of spectra, it is clear that we need to
make use of all physical a priori constraints available to improve and stabilize the fit. If
this is not enough, we may also introduce constraints corresponding to approximations.
One strength of the ‘fit/spec’ software is that it is easy to set up such constraints. In our
example, we will give a detailed description of how the fit of a datacube of an observation
of the active galaxy NGC 2110 was prepared and executed. These observations were
performed in the 630–700 nm wavelength range that included seven important emission
lines (see Figure 4.22).
The first interesting feature provided by ‘fit/spec’ is the use of systems. A system is
a group of lines that share the same velocity and width. This allows a decrease in the
number of degrees of freedom. Lines that can be grouped in a system are lines originating
from the same ion or lines that we expect to originate from the location in the object
(and therefore expect to share the same kinematics). In our example, we know that all
seven lines do not usually display distinct kinematics and we will therefore assume they
can be grouped as a single system. Note that we have to be very cautious with this
type of grouping and some lines like Hβ and [OIII] (not present in our datacube) usually
do not share the same kinematics and should therefore not be included in the same
system.
Second, we need also to introduce constraints between lines in a system. In our example,
! the existence of fixed or bounded line ratios (e.g. between lines of a same doublet);
these constraints will take into account:
! the fact that the width (specified here as the full-width-at-half-maximum, FWHM)
of a line shall be larger than the instrumental one;
Figure 4.23. Configuration file describing the fit to be conducted and the constraints that will
be used. Lines in the system are assigned numbers 1 to 7 (as they appear in the initial description
of the system).
! the fact that the velocities shall be within a few hundreds of km s−1 around the
! the fact that the intensity of the lines shall be positive.

systemic one;
The implementation of these constraints in the software is done through an ASCII con-
figuration file called the ‘stack’ file, which is displayed in Figure 4.23. At this stage, the
fit is performed with a single Gaussian profile.
Initial conditions
We also need to prepare a set of initial conditions for the fit. They will be stored in a
second ASCII configuration file called the ‘start’ file that is a mirror of the ‘stack’ file and
they can be obtained graphically in ‘fit/spec’. Note that in complex objects, it is quite
often necessary to prepare different initial conditions depending on the location within
the object.
Fitting the data

So far, the preparation of the fit has been mainly done manually based on a few spectra
extracted from the whole datacube. A first automated fit of all the spectra (> 1000 in this
case) is then performed and the resulting fits and maps inspected. In our case, despite the
fact that these first maps look good, a check of the fit in the central regions of the object
shows that an additional broad kinematical component is present in the nuclear region
of the galaxy and that it is therefore necessary to describe the line profiles as the sum of
one narrow and one broader Gaussian profile. The ‘stack’ file is modified accordingly (see
Figure 4.24), a new set of initial conditions is prepared and the fit is run again, yielding
much better results. This iterative process is performed until we are satisfied with the fit
(in our case this was enough and the results of the fit can be seen in Ferruit et al., 2004).
Determining the error bars

After all the emphasis in the first chapters on the tracking of the errors during data
processing and analysis, we should not forget that we now need to determine the error
154 Pierre Ferruit
Figure 4.24. Configuration file for a fit with two Gaussian profiles.
bars on the results from the fit. One possible way is to use the error estimate inferred
from the minimization itself (from how ‘sharp’ the minimum toward which the program
has converged is as a function of each fit parameter).
This functionality is not present in the ‘fit/spec’ software and we have therefore used
! Create a synthetic, noise-free spectrum with spectral characteristics representative of

a Monté-Carlo type of approach. The steps were the following:
those of the spectra in the object. Note that it may be necessary to prepare several of
these template spectra if the spectral characteristics of the object vary significantly
! Create a set of noisy spectra with given signal to noise ratios by adding noise to the
within the field of view.
template. For each signal to noise ratio value, a large number of spectra (usually
! Automated fits are performed for all prepared spectra (i.e. 500 per signal to noise
500) are created in order to be able to perform a statistical analysis of the results.
! A statistical analysis is performed on the results to derive the accuracy of the results
value).
as a function of the signal to noise in the input spectrum.

The bottom panels of Figure 4.25 show the results for the velocity measurements. For
this parameter the results are in fact quite straight-forward, but for parameters like the
FWHM, they are usually much trickier to analyse. Note also that this rather simplistic
Monté-Carlo approach does not ‘see’ some contributions to the actual error bars, such
as the errors introduced by the finite accuracy of the wavelength calibration for the
velocity measurements and by the subtraction of the instrumental FWHM for the FWHM
measurements. These additional contributions will have to be determined and added. One
Template (noise-free) spectrum
S/N = 100
1.0
1
500 spectra 0.8
Pixel value (pixel)

0.6
S/N = 1 0.4
0.2
0.0
6300 6400 6500 6600 6700 6800
Position ( )
S/N = 3
S/N = 50
S/N = 5
S/N = 10
Table: fitsn3 Sel: select.and.:ratioSII.1e.10
2350 Table: fitquality
2
2300
1.8
2250
1.6
V_syst1
tmp
2200
1.4
2150 1.2
2100 1
0 0.5 1 1.5 2
logsn (unitless)
0 100 200 300 400 500
Sequence
Figure 4.25. Top panel: sketch illustrating how a Monté-Carlo type of approach is used to
determine the error bars. Bottom panels: fit of the results of Monté-Carlo approach to derive
an analytical relation giving the 1σ error bar on the centroid velocity of a line as a function of
the signal to noise in the input spectrum.
last contribution which is quite often not accounted for because it is extremely difficult
to estimate is the contribution of errors on the stellar continuum subtraction (so-called
template mismatch in particular).
Once all these steps have been performed the results are ready to be included in a
paper.
The all-in-one and ultimate solutions

So far, we have described the processing/analysis of the data as a sequence of various
tasks that are performed independently and with usually a heterogeneous set of tools.
For instruments built around a given, well-defined scientific project, the trend is now to
integrate the analysis tools within the data processing software (the all-in-one solution).
This ensures in particular that all errors are carefully controlled throughout the complete
process.
156 Pierre Ferruit
Table 4.1. Evolution of the total number of ‘pixels’ in a datacube for a set of
arbitrary selected 3D spectrographs.
Instrument Year Number of spaxels Number of spectral pixels Total
TIGER 1987 572 270 154,440

OASIS 1997 1,200 360 432,000
SAURON 1999 1,577 540 851,580
VIMOS 2002 6,400 550 3,520,000
MUSE 2012 90,000 4,096 368,640,000
Last, there are also discussions for some instruments to use an extremely conserva-
tive approach with respect to any resampling of the data and to keep working with
‘detector pixels’ until the very end of the processing/analysis process (the ‘ultimate’
solution). Although very interesting, this approach is extremely complex and requires
the development of specific analysis and visualization tools for every single step of the
data processing/analysis. This has not been implemented to date.
4.9 Conclusions
This course has presented a short but hopefully useful tour of various analysis methods
and strategies. The important message is to never forget to keep track of the noise and
systematic errors throughout the complete data analysis process.
If we make a critical analysis of the various methods that have been presented, one
relatively negative conclusion can be drawn: there are not enough real 3D tools. Most
of the tools presented correspond to generalizations of classical, already-existing tools
to a collection of spectra. They do not make use of the additional spatial information
usually present in the datacube (e.g. the fact that most of the time neighbouring spaxels
are correlated by the spatial PSF) and treat each spectrum independently of the others.
This clearly outlines that a significant effort has still to be made to take full advantage
of the 3D information contained in the datacubes.
Last, looking at how things will probably evolve within the next 5–10 years, there is a
huge challenge in front of us. This is illustrated by the numbers in Table 4.1 that show the
increase in size of the datasets over the years. Many of the analysis steps presented in this
document still rely heavily on a ‘manual’ handling of the data. This will not be possible
with data from, for example, the MUSE instrument. A huge effort will therefore have to
be made to prepare the automated tools that will be needed for the next generation of
3D spectrographs.
REFERENCES
(1999), A&AS, 136, 189
Bender, R. (1990), A&A, 229, 441
Bertin, E., Arnouts, S. (1996), A&AS, 117, 393
Falcón-Barroso, J., Sarzi, M., Bacon, R., Bureau, M., Cappellari, M., Davies,
R.L., Emsellem, E., Fathi, K., Krajnović, D., Kuntschner, H., McDermid, R.M.,
Peletier, R.F., de Zeeuw, T. (2006), New Astronomy Review, 49, 515
Ferruit, P., Mundell, C.G., Nagar, N.M., Emsellem, E., Pécontal, E., Wilson, A.S.,
Schinnerer, E. (2004), MNRAS, 352, 1180
Gibson, S.J., Taylor, A.R., Higgs, L.A., Brunt, C.M., Dewdney, P.E. (2005), ApJ, 626,
214
Hook, R., Fruchter, S. (2000), Astronomical Data Analysis Software and Systems IX,
ASPCS, Vol. 216, edited by Nadine Manset, Christian Veillet, and Dennis Crabtree
‘Euro3D data format, format definition’, http://www.aip.de/Euro3D, Euro3D data format:
Format definition, issue 1.2
Rousset (1992), ‘Contribution des méthodes numériques au dépouillement des données du spec-
trographe intégral de champ TIGER’, PhD thesis, Univ. J. Monnet de Saint-Etienne
Sánchez, S.F. (2004), AN, 325, 167
Wilman, R.J., Gerssen, J., Bower, R.G., Morris, S.L., Bacon, R., de Zeeuw, P.T.,
Davies, R.L. (2005), Natur, 436, 227
5. Science motivation for integral field
spectroscopy and Galactic studies
FRANK EISENHAUER
This lecture introduces the scientific motivation for integral field spectroscopy (IFS)
and describes the results from this novel technique in Galactic studies as of 2005. The
following chapter by Luis Colina then picks up on extragalactic studies and rounds out
the picture giving an outlook to future integral field spectroscopic studies. Following the
! science motivation;
five one-hour lectures, this chapter is broken down into sections on:
! the Galactic Centre black hole;

! the Galactic Centre stellar population;
! star formation; and
! the Solar System.
The publications discussed in this lecture are selected on the basis of their scientific
impact as measured by the citation index and/or because they specifically qualify for a
lecture from a didactic point of view.
5.1 Science motivation

Motivation for the development and application of integral field spectroscopy is manifold
and is closely tied to personal interest and background. This is most obvious in the
history of instrument development, outlined in the beginning of this section. The ‘ideal’
objects for integral field observations can be well characterized by their angular size,
substructure and spectral characteristics. These qualifiers are used to identify the ideal
objects on the distance ladder. The section ends comparing the apparent interest of the
Galactic and extragalactic astronomical communities in integral field spectroscopy.
5.1.1 Motivation for integral field spectroscopy

The development of the first-generation integral field spectrographs was largely driven by
an ‘experimental’ motivation. These instruments were built by small teams that wanted to
invent a new technology. Typical examples for this early epoch are the ARGUS, MPE 3D,
MPFS and Tiger instruments. This is well illustrated by a sentence from Bacon et al.
(1995) on the Tiger instrument: ‘Although it was primarily developed to validate the
lenslet integral field spectroscopy (IFS) concept, it has provided a number of scientific
results on a large variety of astrophysical objects.’
The second-generation integral field spectrographs were then strongly motivated by
the astronomical applications, and these instruments were built by medium-size groups,
partly in international collaborations: typical examples for this present generation of
instruments are: OASIS, SAURON, VIMOS, NIFS, OSIRIS and SINFONI. Many of
the instruments are permanently installed as facility instruments at large observatories.
One finds both instruments for a wide variety of applications and instruments dedicated
to very specific goals. The change from the experimental to the astronomical motiva-
tion is nicely illustrated by another citation from the above authors, now on SAURON,
the successor of Tiger: ‘Its design is optimized for studies of the stellar kinematics, gas
158
Science motivation for IFS and Galactic studies 159
kinematics, and line-strength distributions of nearby early-type galaxies’ (Bacon et al.,
2001a).
The development of a third generation of integral field spectrographs has recently
started. These instruments are built by large international consortia and are usually jus-
tified by a wide variety and mainstream science cases. Examples of this third generation
of instruments are MUSE (approximately 8.4 million euro, 172 staff years, 7 institutes)
and KMOS (approximately 6.5 million euro, 165 staff years, 6 institutes).
Presently, several integral field spectrographs are also being built or developed for
upcoming missions, for example PACS on the Herschel satellite and Near-Infrared Spec-
trograph (NIRSpec) on the James Webb Space Telescope (JWST).
5.1.2 Fundamental considerations based on technical constraints

The parameter space of integral field spectrographs is spanned by the number of spatial
and spectral elements. Although there is no fundamental limit on the number of these
elements, most integral field spectrographs are in practice (still) limited by the number
of available detector pixels, and the fact that array detectors are more or less squarish.
Therefore it turns out that the various integral field spectrographs are quite comparable
in terms of number of spatial and spectral elements. To estimate the typical numbers, we
assume a detector with 2k × 2k pixels. The number of spectra is within a factor of a few
equal to the number of rows, i.e. approximately 2000. This is obvious for mirror-slicer inte-
gral field units (see Bershady’s chapter 3 on instrumentation), but also most lenslet inte-
gral field units turn out to have a comparable number of spectra because the gain by stack-
ing several spectra in one column is compensated by the gaps between the spectra. Assum-
ing
√ a square √ field of view, the typical integral field spectrograph then covers of the order
4000 × 4000 ≈ 50 × 50 pixels. The number of pixels for every spectrum is limited by
the number of pixels per column. Mirror-slicer integral field units take full advantage of
them; lenslet IFUs typically stack 2–4 spectra per column. Therefore the typical num-
ber of spectral elements is about 1000. For comparison, the typical number of elements
(spatial × spatial × spectral) for a long-slit spectrograph are 1 × 1000 × 1000, for an
echelle spectrograph 1 × 1 × 1,000,000 and for an Fabry–Perot spectral imager 1000 ×
1000 × 1.
The size of each spatial pixel on the sky, and therefore the total field of view, is also
quite constrained: the argument for the lower limit is simply the diffraction limit of the
telescope, which is approximately 20 mas for a 10 m telescope at a wavelength of 1 µm.
Therefore, the total field of view of a diffraction-limited integral field spectrograph is
of the order 50 · 20 mas × 50 · 20 mas = 1′′ × 1′′ . The maximum size of a pixel on the
sky can be derived from the Lagrange invariant in optics, the small physical size of a
detector pixels, and the practical limitation on building a camera with an f -number faster
than 1. The Lagrange invariant implies that the étendue = ‘area-solid angle product’
is preserved in imaging optics (and in general can only increase, for example through
the focal-ratio degradation of a fibre). Applied to an integral field spectrograph, this
implies that telescope diameter × pixel size on sky [in rad] = physical pixel size ×
beam opening angle at detector [in rad]. If we assume a typical pixel size of 10 µm,
and again a 10 m telescope, the maximum pixel size on sky for an f /D = 1 camera
(i.e. beam opening angle at detector is approx. 1 rad) is approximately 10−6 rad or 0.2′′ .
Therefore, the maximum field of view of an integral field spectrograph is of the order
50 · 0.2′′ × 50 · 0.2′′ = 10′′ × 10′′ .
160 Frank Eisenhauer
In summary, we find that a typical integral field spectrograph has of the order 50 × 50
spatial pixels with a field of view of 1′′ × 1′′ to 10′′ × 10′′ , and of the order 1000 spectral
pixels.
5.1.3 Ideal and unsuitable objects for integral field spectroscopy

The typical number of spatial and spectral pixels, and the typical field of view are criteria
for ‘ideal’ and ‘unsuitable’ objects for integral field spectroscopy. Clearly unsuitable are
! cover a very large field of view (larger than several square arcminutes);
objects that:
! are sparse and point-like (only few objects per several square arcminutes);
! require observations of a large spectral range at very high spectral resolution (larger
than several 10,000, covering a full atmospheric band).
! (all-sky) surveys, for example the Wisconsin Hα Mapper Northern Sky Survey is
Typical examples for unsuited science cases are:
! multi-object spectroscopy of loose stellar fields or extragalactic planetary nebulae,

done with a Fabry–Perot spectral imager;
done with fibre-fed (e.g. 2dF at the Anglo-Australian Telescope, AAT) or slit-mask
(e.g. Focal Reducer and Low Dispersion Spectrograph, FORS at the Very Large
! high spectral resolution (echelle) spectroscopy for planet search from Doppler velocity
Telescope, VLT) multi-object spectrographs;
measurement (e.g. High Accuracy Radial Velocity Planet Searcher, HARPS at the
European Southern Observatory, ESO, 3.6 m telescope) or the Lyman-alpha forest
observed in quasars (e.g. High Resolution Echelle Spectrometer, HIRES at the Keck
Telescope).
! resolved at the seeing- or diffraction-limited resolution of optical telescopes;

Ideal objects fulfil one or more of the following criteria:
! position is not accurately known;

! no (or unknown) spatial symmetry;
! rich spectrum or unknown redshift.
Examples of ideal objects in our Solar System are the planets, their moons and asteroids.
Ideal Galactic objects are, for example (roughly in the order of distance), brown dwarfs
and low-mass stars, stellar accretion discs, young stellar objects (YSOs), jets, Herbig–
Haro objects, asymptotic giant branch (AGB) stars, proto and young planetary nebulae,
young star clusters, the Galactic Centre region with its black hole, and globular clusters.
Examples of ideal extragalactic objects are (also ordered by distance) the nuclear and old
stellar populations of nearby galaxies, nearby massive black holes, nearby active galactic
nuclei (AGN), nearby starbursts, ultra-luminous infrared galaxies (ULIRGs), quasar host
galaxies, high redshift star-forming galaxies and Lyman-break galaxies.
The science case for the above examples is described in detail in various instrument
proposals and publications, for example in ‘Science with NIFS, Australia’s first Gemini
Instrument’ (McGregor et al., 2001) and ‘SINFONI a High Resolution Near-Infrared
Imaging Spectrometer for the VLT’ (Genzel et al., 1997).
5.1.4 Examples of integral field spectroscopic studies

After the pioneering work with the ‘first generation’ principal investigator instruments,
the race began with the public availability of facility instruments on large telescopes. This
section highlights a number of refereed publications from recent years based on integral
field spectroscopic observations.
As outlined above, the ideal objects are spatially resolved at the diffraction- to the
seeing-limit of optical telescopes, i.e. those objects having structure on the 0.02–0.2′′
scale and cover a field of view of approximately 1–10 . Accordingly, the physical size
′′
and substructure of ideal objects scale with distance. Therefore, the following list is
ordered by the distance, climbing the distance ladder by typically a factor 10 with each
rung.
The typical distance scale for the Solar System is one astronomical unit. An angular
separation of 1′′ corresponds to approximately 700 km, the ideal (1–10′′ ) field of view
(FoV) of an integral field spectrograph corresponds to 700–7000 km, e.g. the size of moons,
continents of planets and asteroids. The angular resolution (0.02–0.2′′ ) corresponds to
10–100 km, e.g. big volcanoes. Although many Solar System objects have already been
visited by spacecraft, there is still need for ground-based observations, for example to
obtain high-resolution spectra, and for monitoring long-term trends and seasons. Exam-
ples of Solar System studies taking advantage of integral field spectroscopy are ‘Martian
surface mineralogy from 0.8 to 1.05 µm TIGER spectra–imagery measurements in Terra
Sirenum and Tharsis Montes formation’ (Martin et al., 1996), ‘The three-dimensional dis-
tribution of Titan haze from near-infrared (NIR) integral field spectroscopy’ (Ádámkovics
et al., 2006) and ‘Near-Infrared Compositional Mapping of Vesta’ (Dumas et al., in
preparation).
The next step on the distance ladder leads to the most nearby stars at about 1 pc
(1′′ = 1 AU), for example Proxima Centauri. At that distance the ‘typical’ FoV covers
1–10 AU, therefore matching the size of planetary systems. The resolution is about 0.02–
0.2 AU, comparable to the separation of planets. However, present-day adaptive optics
and instruments do not provide enough contrast and no publications have emerged from
such observations yet.
The most nearby star-forming regions – for example the Taurus molecular cloud –
are located at a distance of about 100 pc (1′′ = 100 AU). The FoV of 100–1000 AU is
comparable to the size of circumstellar discs, and the angular resolution is 2–20 AU,
comparable to the orbital radius of Jupiter. So far, publications have concentrated on
prototypical stars, jets and their interaction with the ambient interstellar medium: ‘A
near-infrared spectral imaging study of T Tau’ (Herbst et al., 1996; Kasper et al., 2002),
‘Sub-arcsecond morphology and kinematics of the DG Tauri jet in the [OI]λ6300 line’
(Lavalley et al., 1997) and ‘The three-dimensional structure of HH 32 from GMOS IFU
spectroscopy’ (Beck et al., 2004).
The closest high-mass star formation regions, for example the Orion Nebula, are at a
distance of about 1 kpc (1′′ = 1000 AU). The FoV of 1000–10,000 AU and the resolution
of 20–200 AU is still appropriate to cover and resolve circumstellar discs and jets. The
publications address the processes involved in high-mass star formation, for example
‘Collimated molecular jets from high-mass young stars: IRAS 18151-1208’ (Davis et al.,
2004), and the impact of winds and radiation from the high-mass stars on circumstellar
discs of nearby stars, for example in ‘GEMINI multi-object spectrograph integral field
unit spectroscopy of the 167-317 (LV2) proplyd in Orion’ (Vasconcelos et al., 2005).
The Galactic Centre is located at a distance of about 10 kpc (10′′ ≈ 0.5 pc). At that
distance the FoV of 0.05–0.5 pc corresponds to the core size of a typical star cluster
and the resolution of 200–2000 AU is close to a light-day. Being a unique laboratory for
studying supermassive black holes and their environment, the Galactic Centre has been
the subject of a larger number of publications from integral field spectroscopy, for example
on ‘The nuclear cluster of the Milky Way: star formation and velocity dispersion in the
central 0.5 parsec’ (Krabbe et al., 1995) and ‘SINFONI in the Galactic Centre: young
stars and infrared flares in the central light month’ (Eisenhauer et al., 2005). A distance
of 10 kpc is also the typical scale for our Galaxy. Consequently, integral field observations
cover a wide variety of objects at this distance. Examples are ‘Integral field spectroscopy
of faint halos of planetary nebulae’ (Monreal-Ibero et al., 2005) and ‘The X-ray binary
X2127+119 in M15: evidence for a very low-mass, stripped giant companion’ (Van Zyl
et al., 2004).
At a distance of 1 Mpc (1′′ ≈ 5 pc) we get to the closest disc galaxies, for example the
Andromeda galaxy M31. The typical FoV covers 5–50 pc, e.g. nuclear starburst regions,
and the spatial resolution is about 0.1–1 pc, the typical separation of stars in spiral arms.
Publications have so far concentrated on the nuclear regions, for example in ‘The M 31
double nucleus probed with OASIS. A natural vec m = 1 mode?’ (Bacon et al., 2001b).
The closest clusters of Galaxies – for example the Virgo cluster – are located at a
distance of about 10 Mpc (1′′ ≈ 50 pc). The FoV corresponds to about 50–500 pc, and
the resolution of 1–10 pc is still good enough to resolve individual starburst regions.
Again, publications focus mainly on the dynamics and properties of nuclear regions,
for example ‘The SAURON project: III. Integral-field absorption-line kinematics of 48
elliptical and lenticular galaxies’ (Emsellem et al., 2004), and ‘SINFONI adaptive optics
integral field spectroscopy of the Circinus galaxy’ (Müller-Sánchez et al., 2006).
Yet another factor 10 more distant (about 100 Mpc, 1′′ ≈ 0.5 kpc) are massive galaxy
clusters – for example the Coma Cluster – which host ‘extreme’ galaxies like ultra-
luminous infrared galaxies (ULIRGs). The field of view at that distance is approximately
0.5–5 kpc, about the size of galactic bulges. The resolution is about 10–100 pc. This is
sufficient to spatially resolve the velocity field of merging galaxies, a published example
is ‘Stellar dynamics and the implications on the merger evolution in NGC 6240’ (Tecza
et al., 2000).
The most distant objects observed so far with integral field spectrographs are at red-
shifts of z ≈ few (1′′ ≈ 5 kpc at z = 1). This early epoch – several Gyr after the Big
Bang – is particularly interesting because of its peak in cosmic star formation. The FoV
of about 5–50 kpc (at z = 1) covers one or more galaxies. The resolution of 0.1–1 kpc
allows us to resolve regions with sizes comparable to galactic bulges. The publications
focus on characterizing the physical properties of various galaxy samples, for example
‘Optical and near-infrared integral field spectroscopy of the SCUBA galaxy N2 850.4’
(Swinbank et al., 2005) and ‘SINFONI integral field spectroscopy of z ≈ 2 UV-selected
galaxies: rotation curves and metallicity gradients’ (Förster Schreiber et al., 2006), but
also cover various other topics like the feedback mechanism to the intergalactic medium,
for example in ‘Deep SAURON spectral imaging of the diffuse Lyman α halo LAB1 in
SSA 22’ (Bower et al., 2004).
The above examples illustrate the start of the race, and many more exciting results
are expected in the coming years. In parallel, several groups have started with the next-
generation integral field spectrographs and telescopes. The outlook for this future science
is given in the lecture by Luis Colina.
5.1.5 Galactic versus extragalactic science

The science cases outlined in the previous section cover Galactic and extragalactic objects
equally. However, recent conferences indicate that the major science driver is extragalac-
tic. This can be seen from the rough fraction of Galactic versus extragalactic proceedings
! ‘Extragalactic science 3D optical spectroscopic Methods in Astrophysics’, Marseilles

articles in the last five big conferences on science with integral field spectrographs:
! ‘Imaging the universe in 3D’, Walnut Creek 1999 (Van Breugel and Bland-Hawthorn,
1994 (Comte and Marcelin, 1995): 40% Galactic versus 60% extragalactic;
! ‘EURO 3D science conference’, Cambridge 2003 (Astronomische Nachrichten,

2000): 10% Galactic versus 90% extragalactic;
325(2)): 10% Galactic versus 90% extragalactic;

! ‘Workshop on AO assisted IFS’, La Palma 2005 (Rutten et al., 2006): 40% Galactic
! ‘Science perspectives for 3D spectroscopy Conference’, Garching 2005 (Kissler-Patig

versus 60% Extragalactic;
et al., 2006): 20% Galactic versus 80% extragalactic.

In summary, about 25% of the proceedings articles of these conferences are on Galactic
science, and 75% on extragalactic science. A similar trend can be seen at this IAC Winter
School, for which only less than 20% of the astronomical poster contributions are on
Galactic topics, and more than 80% on extragalactic topics.
5.1.6 Overview of Galactic science

The following four sections give an overview of Galactic science with integral field spectro-
graphs as of 2005. All subsequent citation counts are up to late 2005. The most successful
application of integral field spectroscopy to Galactic science up to 2005 is Galactic Centre
research with about 600 citations, followed by star formation research with about 125
citations. Solar System studies and other topics have had limited impact so far and only
! On the Galactic Centre: Eisenhauer et al. (2003a), ≈ 71; Eisenhauer et al. (2005),
a few citations. For the papers discussed in this lecture we find:
≈ 16; Genzel et al. (1996), ≈ 130; Genzel et al. (2000), ≈ 120; Genzel et al. (2003a),
! Related to star formation: Beck et al. (2004), ≈ 3; Davis et al. (2004), ≈ 7; Dougados
≈ 89; Krabbe et al. (1995), ≈ 178;
et al. (2003), ≈ 1; Herbst et al. (1996), ≈ 38; Kasper et al. (2002), ≈ 6; Lavalley et al.
! On Solar System objects: Ádámkovics et al. (2006); Martin et al. (1996), ≈ 5.

(1997), ≈ 34; Lavalley-Fouquet et al. (2000), ≈ 35; Vasconcelos et al. (2005), ≈ 0;
However, it should be noted that this list is to a large extent limited to the ‘pioneering’
phase of IFS, and that the real impact is just about to start with the installation of
several facility instruments at large observatories.
5.2 The black hole in the Galactic Centre

This section is about the black hole at the Galactic Centre and how IFS contributes to
prove its existence. The subsections introduce the subject of supermassive black holes,
give the line of argument for proving the existence of the black hole, describe the mea-
surement of its mass and distance, and discuss the paradox of its faintness. Many of the
publications also address the stellar population close to the Galactic Centre. For didactic
reasons, however, this topic is discussed in a separate section.
Excursion: history of supermassive black holes

Presently, we know two kinds of black holes: stellar black holes with a mass of ≈ 10–
30 M⊙ and supermassive black holes with a mass of ≈ 105 –109 M⊙ . While the formation
of a stellar black hole from the collapse of a massive star in a Type II supernova is well
understood, and pulsar binary systems provide a variety of good evidence for the black
hole hypothesis, the case for supermassive black holes has been less clear for a long time.
The first strong evidence for such extreme mass concentrations was the discovery of the
quasar 3C273 by Schmidt (1963), who even made it to the cover page of Time magazine.
Based on the observed Doppler shift, this 13th magnitude ‘star’ moves away from us at
16% of the speed of light. If this redshift is correctly interpreted in terms of the Hubble
flow, the object is located at a distance of approximately 1 Gpc, corresponding to a
distance modulus of ≈ 40. Therefore the absolute magnitude of 3C273 must be about
−27 mag, and its luminosity about 1013 L⊙ . As the brightness of the object varies over a
period of days, its size must be smaller than light-days, i.e. of the order less than 1000 AU.
Figure 5.1. Adaptive optics K-band (1.95–2.45 µm) image of the Galactic Centre region. The
field of view is approximately 20′′ × 20′′ . The arrows indicate the position of the black hole, which
is not visible in this image. Only during sporadic flares (every couple of hours) is the emission
strong enough to be detected. The angular resolution in the image is approximately 50 mas. For
comparison, the Schwarzschild radius of the Galactic Centre black hole is approximately 10 µas.
In other words, 3C273 emits the light of the whole Galaxy from a region comparable to
the size of the Solar System.
Subsequent observations of radio halos of quasars indicated that the energy source is
gravitation and not nuclear fusion: the energy contained in such radio halos is up to about
1054 J. If nuclear fusion were the energy source – for which the efficiency of converting
mass to radiation is of the order 1% – about 109 M⊙ would be required to power the radio
halos. However, the self-gravitational energy of this mass when squeezed into a couple of
light-days would exceed 1055 J, more than 10 times the fusion energy. Therefore, most of
the energy released in quasars must have a gravitational origin.
The evidence that the light source in quasars is as small as a black hole, however, was
weaker and more indirect: it was based on observations of radio jets. These jets reach
far out into intergalactic space and as such must ‘remember’ their ejection direction for
quite a long time. The orientation of the emitting region must thus be very stable, like a
‘gyroscope’. This indicates that the emitting object is probably spinning very fast. The
other key observation was that radio jets move close to the speed of light. This implies
indirectly that the speed of the spinning material is also close to that of light, i.e. it must
originate close to the Schwarzschild horizon of a black hole.
To summarize the epoch of 1970–90, most astronomers believed the supermassive black
hole picture. They found an enormous amount of indirect evidence, i.e. no other model
could reproduce the measurements equally well. But there has been no strict proof that
supermassive black holes exist, i.e. that the mass is enclosed within the Schwarzschild
radius, and that there is an event horizon. So the search for supermassive black holes
became a very hot topic, both in external galaxies (see lecture by Luis Colina) and in
our Milky Way (this lecture).
Excursion: line of argument for proving the Galactic Centre black hole
The line of arguments to prove the existence of the supermassive black hole in the Galactic
Centre is outlined in Figure 5.2. The first step is to measure its mass at ever-smaller radii.
Down to a few thousand Schwarzschild radii (RS ) this is accomplished by measuring
the orbits of stars with high angular resolution imaging (Figure 5.1) and integral field
spectroscopy (Figure 5.5). The combined astrometric and radial velocity observations also
provide the best measurement of the distance to the Galactic Centre, a prerequisite for
Mass Radius
Velocity and
3 x 10 M Few thousand R
orbits of stars
Non-motion of Size of
radio source versus Few ten R radio
speed of stars source
Brightness variations Few R

and spectral changes
Faintness, SED
Event horizon
Figure 5.2. Line of arguments to prove the existence of the supermassive black hole in the
Galactic Centre.
pinning down the absolute mass and radius. Already, these measurements exclude most
alternatives (for example a very dense cusp of stars or stellar remnants) to the black
hole hypothesis. Radio measurements of the size of Sgr A* (the name of the radio source
which is supposed to harbour the black hole) and the fact that the radio source hardly
moves when compared to the stars then make the case that the mass is indeed within
just a few tens of RS . The observed brightness variations and changes in the slope of the
spectra are interpreted to arise partly from orbital motions as close as a few RS . Finally,
it is the faintness of the Galactic Centre and the detailed shape of its spectral energy
distribution (Figure 5.6) that provide the theoretical argument that there is indeed a
black hole with a Schwarzschild horizon.
5.2.1 The nuclear cluster of the Milky Way: star formation and velocity dispersion
in the central 0.5 parsec
The first integral field spectroscopy observations of the Galactic Centre were published
by Krabbe et al. (1995). Their observations show that the radial velocity dispersion of 35
early- and late-type stars with distances of 1–12′′ from Sgr A* is 154 ± 19 km s−1 , which
strongly favours the presence of a central dark mass of 3 × 106 M⊙ within 0.14 pc of the
dynamic centre.
The observations were carried out in 1994 with the world’s first infrared integral field
spectrograph MPE 3D (Weitzel et al., 1996) visiting the ESO-MPI 2.2 m telescope in
La Silla. The spectra cover the wavelength range 1.9–2.4 µm at a spectral resolution
λ/δλ ≈ 1000. In total, 12 early-type emission line stars and 9 late-type stars with CO
band heads (see excursion) could be identified in the central 5′′ radius.
Excursion: Stellar spectra of early- and late-type stars in the infrared

Near-infrared spectra of stars have characteristic emission and absorption lines that allow
accurate spectral classification. In particular, it is very easy to distinguish between early-
type emission line stars and late-type stars with CO band-head absorption. Figure 5.3
shows three stars from the Galactic Centre region with typical spectra. Early-type emis-
sion line stars are young massive (more than about 20 M⊙ ) stars, for example Wolf–Rayet
and Of stars (f indicating emission). They show strong emission from one or several lines
of hydrogen, helium I and II, and ionized nitrogen and carbon, which originate mostly
in stellar winds. Late-type stars – for example, K and M giants – do not show emission
lines, but have deep CO band-head absorption from the outer parts of their cool (less
than about 4500 K) atmospheres.
Figure 5.3. Typical spectra of early-type emission line stars (left) and late-type stars (right).
The figure shows three stars from the Galactic Centre region (courtesy of Genzel et al., 1996).
The emission lines from the early-type stars and the CO band-heads of late-type stars
allow the measurement of the radial velocity of the stars from the Doppler shift. The
measured radial velocities and projected distances from the Galactic Centre then give
the enclosed mass. From the virial theorem (see excursion) the estimate for the enclosed
mass within 0.14 pc is 4.0 × 106 M⊙ . The Bahcall and Tremaine (1981) estimator (see
excursion) for the enclosed mass within 0.19 pc is 6.3 × 106 M⊙ . When adding the veloci-
ties from the more distant stars, and also considering the gas dynamics at larger radii, the
gravitational potential in the Galactic Centre can be well fitted with the combination of
a 3 × 106 M⊙ central (black hole) mass and an isothermal stellar cluster with 3 × 105 M⊙
within a core radius of 0.5 pc (Figure 5.4). However, the data are still consistent with
a central mass dominated by a very compact cluster of stellar remnants (e.g. about 105
stellar black holes with a typical mass of 10 M⊙ ).
Excursion: from velocity to mass I: the virial theorem

The measurement of radial velocities and projected distances allow the determination of
the central mass from the virial theorem. The proof is well known, and is repeated here
(following Bahcall and Tremaine, 1981) to give the basis for the various mass estimates
used in the following papers. Assuming that the gravitational potential in the centre of
the Galaxy is dominated by a point mass, the stars can be described as test particles,
which follow the gravitational law:
GM r
r̈ = − , (5.1)
r3
Figure 5.4. Mass of the Galactic Centre black hole: the plots show the various estimates for
the encircled mass profile from the first integral field spectroscopic observations in 1995 to the
state of the art in 2005 (courtesy of Genzel et al., 1996; 2000; Krabbe et al., 1995).
where r is the vector from the point mass to the star, and G the gravitational constant.
If we then write
) *
d2 1
r · r = r · r̈ + ṙ · ṙ (5.2)
dt2 2
and use r · r = r2 , ṙ · ṙ = v 2 , and average <> over all stars, we get
1 d2 1
< r2 > = −GM < > + < v 2 > . (5.3)
2 dt2 r
Assuming that the ensemble of stars is in a statistical steady state, < r2 > is independent
d2 2
of time and dt 2 < r > = 0. If we further assume that the system is spherically symmetric,
then < v 2 > = 3 < vz2 > and < 1r > = π2 < R1 >, where vz is the radial velocity and R
the projected distance to the centre. From that, we get the virial estimator for the central
point mass:
3π < vz2 >
M= . (5.4)
2G < R1 >
Excursion: from velocities to mass II: Bahcall–Tremaine estimator

As outlined by Bahcall and Tremaine (1981), the virial theorem estimator has some
serious shortcomings: it is biased in the sense that the virial mass is not necessarily equal
to the real mass when derived from only a finite number of stars. Second, the virial
theorem estimator is inefficient, i.e. the variance in the estimate is larger than necessary,
because it weights objects at small distances too strongly. The problem is inherent to
the virial theorem estimator and can be avoided by using an estimator based on the
projected mass, q:
vz R
q= . (5.5)
G
This projected mass treats all test particles at all distances on an equal information basis.
Doing some mathematics to derive the ensemble average < q > for an arbitrary distri-
bution of stars, Bahcall and Tremaine (1981) obtain the relation with the eccentricities
of the stellar orbits:
πM / 0
<q>= 3 − 2 < e2 > . (5.6)
32
If all the stars are on circular or radial orbits, then < e2 > = 0 or 1. For isotropic
orbits < e2 > = 12 . In the absence of any information on the eccentricities, the authors
recommend using
24
M= < vz2 R > . (5.7)
πG
Other than the virial estimate, this Bahcall–Tremaine estimator is not biased when
applied to only a small number of stars and also weights the test particles properly.
The disadvantage is its dependence on the eccentricity distribution.
5.2.2 The dark mass concentration in the central parsec of the Milky Way
In this second paper, Genzel et al. (1996) present an extended dataset, and improve their
mass estimate by considering a rotation component in the measured velocity dispersion
and by simultaneously fitting the observed surface density of stars.
The field of view covered in the observations is approximately 20′′ × 20′′ . Taking advan-
tage of additional high angular resolution speckle imaging data, the authors could extract
the spectra of 198 late-type stars, and 25 early-type stars. The projected stellar velocity
dispersion increases statistically very significantly from about 55 km s−1 at a projected
distance of 5 pc, to 180 km s−1 at 0.1 pc. When plotting the measured radial velocities
against the declination, a significant rotation of the stellar population is obvious. The
late-type stars follow general Galactic rotation, while the early-type stars show counter-
rotation. One therefore has to subtract (in quadrature) the rotation component
2vr2ot (R)
Mr ot = <R> (5.8)
πG
from the mass as derived from the virial theorem or the Bahcall–Tremaine estimator (see
previous section). vr ot (R) is the best-fit rotation curve.
The shortcoming of the virial theorem and Bahcall–Tremaine estimator is obvious: their
application assumes implicitly a homogeneous distribution of stars, but the measured
distribution is by no means homogeneous.
Therefore, Genzel et al. (1996) have also derived a less-biased enclosed mass esti-
mate from the first moment of the collisional Boltzmann equation, the ‘Jeans equation’
(for details see Binney and Tremaine, 1987). The problem is that the formula contains
the three-dimensional distance and the space density, and not the measured projected
distance and surface density. The transformation from volume to surface quantities is
mathematically described by Abel integrals. Since the deconvolution of the projected
quantities to the corresponding volume quantities is very sensitive to individual data
points and their errors, the authors parameterize the velocity dispersion and density dis-
tribution and find the best-fitting model to the data. The analysis gives a central dark
6
mass of 2.5–3.2 × 10 M⊙ at 6–8 σ significance (Figure 5.4). The data also show that
the mass-to-light ratio significantly increases towards the Galactic Centre up to at least
100 L ⊙,M2 ⊙µm , indicating the ‘dark’ nature of the central mass. It is therefore either a com-
pact cluster of 105 stellar black holes with a typical mass of 10–20 M⊙ , or a single massive
black hole.
5.2.3 Stellar dynamics in the Galactic Centre: proper motions and anisotropy
This paper by Genzel et al. (2000) combines proper motion data from speckle interfer-
ometric imaging with radial velocity measurements from integral field spectroscopy in
a new analysis of the stellar dynamics in the Galactic Centre. They find that the sky-
projected velocity components of the young, early-type stars indicate significant devia-
tions from isotropy, with a strong radial dependence. Most of the bright emission line
stars at separations from 1′′ to 10′′ from Sgr A* are on tangential orbits. This tangential
anisotropy of the He I stars and most of the brighter members of the IRS16 complex
is largely caused by a clockwise (on the sky) and counter-rotating (line of sight, com-
pared to the Galaxy), coherent rotation pattern. The assumption of velocity isotropy
in the above mass estimates is clearly violated. The authors thus explicitly include the
velocity anisotropy in estimating the central mass distribution. This Leonard and Mer-
ritt (1989) mass estimator (see excursion) confirms previous conclusions that a compact
central mass concentration (central density ≥ 1012.6 M⊙ / pc3 ) is present and dominates
the potential at 0.01–1 pc. Depending on the modelling method used, the derived central
mass is 2.6–3.3 × 106 M⊙ .
Excursion: from velocity to mass III: the Leonard–Merrit estimator

The virial theorem and Bahcall and Tremaine mass estimator assume an isotropic dis-
tribution of stellar orbits. To overcome this limitation, Leonard and Merrit (1989)
introduced – assuming a spherically symmetric, non-rotating system – an isotropy-
independent mass estimator, which, however, requires data on all three space velocities:
M = < r(2σr2 + 2σt2 ) >, (5.9)
where σr and σt denote the velocity dispersion parallel and perpendicular to the radius
vector, r. When transformed to the projected quantities – as observed – the Leonard–
Merrit mass estimator can be written as:
16 / 0
M= 2 < RσR2 > + < RσT2 > , (5.10)
3π
where σR is the line-of-sight velocity dispersion and σT is the dispersion measured in the
velocity perpendicular to the projected radius vector, R.
Excursion: from accelerations and orbits to mass

All mass estimates for the Galactic Centre black hole presented above rely to some extent
on assumptions concerning the distribution function of the stellar orbits, for example
spherical symmetry or isotropy. The combination of radial velocity measurements with
proper motions and the use of advanced mass estimators removed only part of this bias.
The next step forward in the mass determination was the measurement of accelerations
and stellar orbits of close stars (in particular ‘S2’; see Figure 5.5) from multi-epoch
speckle and adaptive optics imaging (Schödel et al., 2002; Ghez et al., 2003). With the
exception of the ambiguity in the inclination, these astrometric observations constrain all
orbital parameters. However, the conversion of the angular semi-major axis – in units of
arcseconds – to a physical distance – in units of metres – and therefore the determination
Figure 5.5. Three-dimensional orbits and distance determination for the Galactic Centre: fit-
ting simultaneously the measured radial velocities and motion on the sky gives directly the
distance to the Galactic Centre. The radial velocity also removes the ambiguity in the inclina-
tion when fitting the stellar orbits solely from their motion on the sky (courtesy of Eisenhauer
et al., 2005).
of the central point mass requires the distance to the Galactic Centre. This distance
is known from various other measurements to be 8.0 kpc, but only to an accuracy of
±0.5 kpc. Since the mass estimate scales with the third power of distance, the distance
is also the major uncertainty (about 20%) in the mass determination.
5.2.4 A geometric determination of the distance to the Galactic Centre

This paper by Eisenhauer et al. (2003a) was the first to report a geometric determination
of the distance to the Galactic Centre. The distance of 7.94 ± 0.42 kpc is derived from
combining the astrometric data of the ‘S2’ star with radial velocity measurements from
adaptive optics long-slit spectroscopy and new integral field spectroscopy.
The idea behind this distance estimate is quite simple and has been known and applied
for decades to binary stars: from Kepler’s law one can calculate all three space velocities.
While the measured radial velocities are independent of distance, the orbit as observed on
the sky (in angular units) is inversely proportional to the distance. Fitting the observed
radial velocities and the orbits on the sky simultaneously thus directly gives the distance
to the Galactic Centre (see also Figure 5.5). The derived distance of 7.94 ± 0.42 kpc was
not only the most accurate primary distance measurement in 2003 but also has minimal
systematic uncertainties of astrophysical origin.
5.2.5 SINFONI in the Galactic Centre: young stars and infrared flares
in the central light month
This paper by Eisenhauer et al. (2005) focuses on the stellar population within the central
light-month and flares from the black hole (see below), but also gives an update on the
mass and distance to the Galactic Centre. The paper is based on the first adaptive optics-
assisted integral field spectroscopy observations of the Galactic Centre with SINFONI
(Eisenhauer et al., 2003b; Bonnet et al., 2004), which allowed measurement of the radial
velocities of about a dozen stars in the central arcsecond, and to add another epoch to
the S2 orbit (Figure 5.5). The updated estimate of the distance to the Galactic Centre
from the S2 orbit fit is 7.62 ± 0.32 kpc, resulting in a central mass value of 3.61 ± 0.32 ×
106 M⊙ (Figure 5.4).
Figure 5.6. Spectral energy distribution (SED) of the Galactic Centre. The measurements of
the slope at near-infrared frequencies strengthen the radiative accretion models and as such
provide further evidence for an event horizon (courtesy of Eisenhauer et al., 2005).
Excursion: Srg A* does not move

As outlined in the line of arguments in Figure 5.2, the stellar orbits measure the enclosed
mass within a few thousand Schwarzschild radii (RS ). The evidence that this mass is
indeed enclosed within a few ten RS comes from radio observations: the argument is
based on two findings; first, that the size of the radio source measured at mm radio
wavelengths is less than about 10 RS ; and, second, that the radio source is indeed the
massive central object. The evidence for the latter is that the radio source does not
move. Very large baseline interferometry gives an upper limit of 0.5 km s−1 perpendicular
to Galactic plane. Such low velocities can be expected only if the radio source cannot be
kicked around by encounters with stars, i.e. is much, much heavier than stars and does
indeed carry the mass as derived from the stellar orbits (Reid and Brunnthaler, 2004).
Excursion: hot spots on the last stable orbit

The evidence that the compact supermassive source is indeed as small as a few RS
comes from observations of infrared flares, which show a modulation with a period of
approximately 17 min (Genzel et al., 2003b). If this modulation reflects the orbital motion
of the emitting region, the short period can only be explained if the ‘hot spot’ moves
close to the last stable orbit, which is at 3 RS for a non-rotating black hole.
5.2.6 Again: SINFONI in the Galactic Centre: young stars and infrared flares
This paper by Eisenhauer et al. (2005) was already mentioned for the mass and distance
determination. The authors also report the first measurements of the slope of the spec-
tral energy distribution (SED) in the infrared during a flare. They find that the SED at
′
1.7–2.45 µm is fitted by a featureless ‘red’ power law Sv ∝ v −α of spectral index −4 ± 1
(Figure 5.6). The observed spectral slope strengthens synchrotron models in which the
infrared emission comes from accelerated, non-thermal, high-energy electrons in a radia-
tively inefficient accretion flow.
This support for the radiatively inefficient accretion is a step towards ‘proving’ the
existence of the event horizon of the Galactic Centre black hole. The main evidence for the
presence of such an event horizon – i.e. that matter, energy and radiation cannot escape
from inside this region – is the faintness of the Galactic Centre supermassive object.
It is several thousand times fainter than what is expected from the mass–luminosity
relation for the Centre of other galaxies (Nagar et al., 2005) and about a million times
smaller than the Eddington luminosity (see excursion). This very low luminosity is best
explained by radiatively inefficient accretion models (see excursion), which inherently
rely on the presence of an event horizon. To ‘prove’ the existence of the event horizon
the observations must constrain the parameters of the accretion models. The strongest
observational constraint for the accretion models – and thus for the presence of an event
horizon – is the SED of the Galactic Centre. In the energy range from radio to X-
rays, the main features of the SED are the synchrotron radiation at radio wavelengths
rising with a power law up to about 1012 Hz, a maximum at sub-millimetre wavelengths,
a strong decline in the infrared, and strong emission at X-ray energies during flares.
The connection between the infrared and X-ray emission is particularly important in
constraining the accretion models – and thus the event horizon hypothesis – because
it is very sensitive to the physical parameters of the emitting region, e.g. size, density,
temperature and magnetic field strength. Coordinated multi-wavelength observations and
infrared integral field spectroscopic observations are targeting this topic.
Excursion: the Eddington luminosity

The Eddington limit is a simple relation for the maximum luminosity of a source with
respect to its mass. It is the limit where the radiation force on the particles in the emitting
region is equal to the gravitational force. While it does not directly apply to emission
from accretion on to a black hole, it is nevertheless often used to characterize the emission
from active galactic nuclei and black holes. In a plasma of electrons and hydrogen atoms,
the radiation force, FR , is dominated by the Thompson scattering of the photons with
the electrons:
σT F
FR = (5.11)
c
) 2 *2
8π e
σT = (5.12)
3 me c2
L
F = (5.13)
4πr2
where σT is the Thompson cross-section for electrons (mass me , charge e, speed of light
c) and F is the radiation flux at the distance r for given luminosity L.
The gravitational force on the plasma from the central black hole is dominated by the
much more massive protons:
GM mP
FG = , (5.14)
r2
where G is the gravitational constant, mP the mass of a proton and M the mass of
the black hole. To maintain a stationary accretion flow in the plasma, gravitation must
dominate over the radiation pressure (FG > FR ). This leads to the Eddington limit:
4πGmP c M
L < LEddington = = 1.3 × 1038 erg s−1 . (5.15)
σT M⊙
Excursion: radiatively inefficient accretion

There are several models to explain the faintness of the Galactic Centre black hole.
A recent summary of the theoretical concepts can be found, for example, in Narayan
(2002). In essence, two ingredients are necessary to reduce the flux by a factor of a
thousand below what is expected from the mass–luminosity relation of active galactic
nuclei. First, one needs a low-density accretion flow that can develop a two-temperature
plasma. Here, the protons are at a low enough temperature (≈ 1010 K) in order not to
contribute significantly to the synchrotron radiation, which is dominated by the higher
temperature (≈ 1012 K) electrons. As the protons are much heavier than the electrons,
this two-temperature, low-density plasma can effectively transfer most of the mass down
to the black hole at low luminosity, and by passing the event horizon permanently trap
the gravitational energy.
5.3 The stellar population in the Galactic Centre

This section is about the stellar population in the Galactic Centre. The subject is closely
related to the black hole, not only because the stars are the ‘test particles’ for deriving
its mass, but also because their formation and evolution is strongly linked to black hole
feeding and growth. Indeed, many papers have already been introduced in the previous
section. This section now focuses on the stellar population in the central parsec, the
‘paradox of youth’ and the stellar dynamics.
Because of the high extinction, deep imaging is limited to near-infrared wavelengths,
and not much colour information can be retrieved for the stars. In addition, the highly
variable extinction makes it very difficult to interpret the colour–magnitude diagram.
Photometric studies cannot even distinguish early- and late-type stars. The need for
spectroscopy was obvious. However, the region is so crowded that classical long-slit spec-
troscopy turned out to be impractical in most cases. In fact the Galactic Centre was a
major scientific driver to build the first near-infrared integral field spectrograph MPE 3D
(Weitzel et al., 1996) and the first adaptive optics-assisted integral field spectrograph
SINFONI (Bonnet et al., 2004; Eisenhauer et al., 2003b) at a 10 m-class telescope.
5.3.1 The nuclear cluster of the Milky Way: star formation and velocity dispersion
in the central 0.5 parsec
This paper, by Krabbe et al. (1995), showed that the central parsec of the Galaxy is
powered by a cluster of about two dozen luminous hot stars.
The authors could identify two main groups of emission line stars and model their
spectra with non-LTE (local thermal equilibrium) stellar atmospheres with winds: the
first group are the so-called ‘He-stars’, which show strong emission in He I and Brγ (see
also Figure 5.3). These stars are classified as WN9 (late nitrogen-rich Wolf–Rayet stars)/
Ofpe (O stars with peculiar emission lines) stars with temperatures of 20,000–30,000 K.
The emission lines show P Cygni profiles, indicating winds with a mass loss rate of
> 10−5 M⊙ yr−1 . Several stars are found to have a very large He I/Brγ ratio and seem
to be pure He-stars. The second group of emission lines stars are of type WC (carbon
>
rich Wolf–Rayet stars) with temperatures ∼ 30,000 K. These stars show C III and N III
emission lines with a width of several 1000 km s−1 , indicating very strong winds. The
progenitor stars must have had zero age main sequence (ZAMS) masses up to about
100 M⊙ . The hot star cluster can fully account for bolometric and ionizing luminosity of
central parsec of the Galaxy.
The authors could also identify more than a dozen late-type stars by the strong CO
band-head absorption in their spectra. They found four supergiants with spectral type
K/M. These stars have K-band luminosities 10,000 times larger than the Sun. There
are also more than 10 late-type giants within the central 8′′ , maybe so-called asymptotic
giant branch (AGB) stars.
Putting together the findings from their integral field spectroscopy, the authors can
model the stellar population and star-forming history (see excursion on stellar popula-
tion modelling). The following observations have to be reproduced by the model: Lb ol ≈
N (KM I)
107 L⊙ , LLLb yo cl ≈ 10, LLLKy c < 0.1, N (OB + WR/HeI with Lb ol > 5.5) ≈ 3, N (OB+W R/HeI) ≈
N (OB)
0.1 and N (W R/HeI) < 0.1.
Even without detailed modelling, some scenarios are obviously excluded; for example,
continuous star formation – i.e. constant star formation over more than Gyrs – cannot
account for the high number of Wolf–Rayet stars. If one uses our Galaxy as a template
for continuous star formation – where we observe only of the order 100 Wolf–Rayet stars
for a total stellar mass of 2 × 1011 M⊙ – one would expect no (< 0.001) WR star in
the central 10′′ with its < 107 M⊙ , but several such stars are observed. Also, constant
star formation – i.e. star formation that started some time ago and has continued at a
constant rate since then – can be excluded: the large LLLb yo cl > 10 would require that the
star formation has started more than 107 years ago. This would imply that the first late-
N (KM I)
type supergiants had formed, so that N (OB+W R/HeI) > 1. However, the observed ratio is
less than 0.1.
The best-fitting model is a small starburst approximately 7 ± 1 Myr ago, with a decay
time of 3–4 Myr. About 103.5 stars have formed in this small starburst. In this scenario
the Galactic Centre stars are presently in a short-lived, post main sequence wind phase.
The presence of the intermediate mass late-type stars cannot be explained by that
scenario and indicates another star formation event about 108 years ago. Another problem
with the small starburst model is that it predicts about 15 early-type (OB) main sequence
stars, which have not been found in the data of Krabbe et al. (1995). However, this non-
detection is solely based on the sensitivity limits at that time, and these stars have been
found with the next generation integral field spectrograph SINFONI at the VLT.
Excursion: stellar population modelling

Several computer programs are available to simulate the integrated properties of stellar
populations, for example ‘Starburst99’ (Leitherer et al., 1999). In these programs the
stellar population is parameterized by the stellar initial mass function (IMF, typically
described by a broken power law with upper and lower mass cut-off), star formation
history (e.g. the age of the burst and its decay time) and stellar evolution and atmo-
spheres (e.g. depending on metallicity). As a result the programs return, for example,
luminosities, number counts and line ratios, which can then be compared to observations.
5.3.2 The dark mass concentration in the central parsec of the Milky Way
In this second paper, Genzel et al. (1996) also analyse in detail the spatial distribution
and dynamics of the central star cluster. In particular, they find that the early-type stars
(He stars, WR/Of stars) are strongly concentrated within a radius of 12′′ , and that these
stars show a large and coherent counter-rotation. This strongly supports their origin in a
well-defined epoch of star formation (see above), presumably when a dense cloud fell into
the central parsec less than 10 Myr ago. Apart from the early-type stars, the late-type
stars (supergiants and bright AGB stars) form a ring at 4–10′′ . The authors find that
the most likely explanation for the absence of the largest red stars (with a stellar radius
of ≈ 50–900 R⊙ ) in the ≈ 5′′ central ‘hole’ is their destruction by collisions with main
sequence, solar-type stars.
Excursion: The paradox of youth

The large number of young (7 Myr), massive (> 10 M⊙ ) stars identified by Krabbe et al.
(1995) and Genzel et al. (1996) so close (< 1 pc) to the central black hole is puzzling,
because all ‘simple’ explanations seem to fail. This puzzle got even more severe when
Ghez et al. (2003) could spectroscopically show that the star S2 orbiting the central
black hole with a semi-major axis of only ≈ 4 mpc is an early-type star, which led them
to introduce the term ‘paradox of youth’. The reason why ‘these young, massive stars
have no right to exist’ is that the strong tidal field from the central black hole prevents
in situ formation, and that the life-time of the massive stars is too short to bring them
from further out at a high enough rate. In somewhat more detail:
The in situ formation of stars from molecular clouds at a distance of 10′′ would require
gas densities of > 109.5 hydrogen atoms cm−3 . However, the densest clouds within 30–60′′
have a factor 1000 smaller densities (≈ 106 hydrogen atoms cm−3 ).
In the migration scenario, where the stars form further out, dynamical relaxation can
slowly bring the stars to the centre. The relaxation timescale for the Galactic Centre
depends on the mass, M , of the stars and is of the order 1500 M yr
M [M ⊙ ] . For comparison, the
main sequence lifetime of a star is of the order 20,000 M yr
M [M ⊙ ] 2 . In order to get the stars down
to the Galactic Centre, the life time has to be larger than the relaxation time, which is
fulfilled only for stars < 1.3 M⊙ . However, the observed stars are much more massive.
Another scenario is the infall of a young cluster, which undergoes core collapse and
loses its outer stars by dynamical friction close to the black hole. The timescale for
dynamical friction is inversely proportional to the cluster mass, and can therefore be as
short as few Myrs. The required cluster mass is of the order > 105 M⊙ at a distance of
5 pc. To avoid tidal disruption before the cluster can core-collapse, the density must be of
the order 107 M⊙ /pc3 and its diameter < 0.4 pc. But there are at present no such clusters
nearby. The most massive clusters Arches and Quintuplet are at about 30 pc and have
only few 104 M⊙ .
Another scenario proposed is the collision of lower-mass stars, which would rejuvenate
the stars’ atmospheres and produce ‘super-blue stragglers’. The collision timescale is of
[ ′′ ] 2
the order 1000 Myr × MR[M ⊙]
. For the 10 Myr young stars this scenario could thus only be
plausible for the central 0.1′′ , but these stars are observed up to 10′′ .
The rotation of the early-type stars discovered by Genzel et al. (1996) gives a hint on
how to resolve the ‘paradox of youth’, supporting star formation from dense molecular
clouds falling into the Galactic Centre. This scenario is supported by the fact that the
stars are counter-rotating the Galaxy, because counter-rotating gas streams have a higher
probability of colliding with ‘normal’ rotating clouds, so that they can lose their angular
momentum and sink to the centre.
5.3.3 Stellar dynamics in the Galactic Centre: proper motions and anisotropy
As described in the discussion on the black hole mass, Genzel et al. (2000) find that many
of the emission line stars at 1–10′′ radius not only show rotation in the radial velocity but
also rotate clockwise on the sky. This supports even more strongly the notion that the
rotation of the young star cluster may be a remnant of the original angular momentum
Figure 5.7. Two stellar discs in the Galactic Centre. Left: measured velocities in the clockwise
disc and perpendicular to it (courtesy of Levin and Beloborodov, 2003). Right: the same for the
counter-rotating disc (courtesy of Genzel et al., 2003a).
pattern in the interstellar cloud from which these stars were formed. The authors also
report the first proper motions of the ‘S-stars’ within the central 1′′ . The absence of CO
emission in the integrated spectrum of that region indicates that these S-stars are also
early-type stars. If they are on the main sequence, they would be of type B0 to B2. As
most of the S-star cluster members are also on clockwise orbits, the authors propose that
the S-stars are those members of the early-type cluster that happen to have low angular
momentum and can thus plunge into the immediate vicinity of Sgr A*.
5.3.4 Stellar disk in the Galactic Centre: a remnant of a dense accretion disk? and
the stellar cusp around the supermassive black hole in the Galactic Centre
Levin and Beloborodov (2003) and Genzel et al. (2003a) were the first to interpret the
observed rotation of the early-type stars in the central 10′′ as a thin stellar disc.
Levin and Beloborodov (2003) find that out of 13 stars whose three-dimensional velo-
cities have been measured by Genzel et al. (2000), 10 lie in a thin disc. If the stars are
rotating in a common disc, the velocity vectors vi of all stars must be perpendicular to
the normal vector, n, of the disc. The authors thus search for the vector n that minimizes
(n · vi )2
N
1 $
χ2 = , (5.16)
N − 1 i=1 (nx σxi )2 + (ny σy i )2 + (nz σz i )2
where nx,y ,z are the components of the normal vector and σxi,y i,z i the errors in the
measured velocity components of the stars. N is the number of stars. Figure 5.7 shows the
velocities of the stars parallel and perpendicular to the disc for the best-fitting orientation.
Simulating an artificial sample of flared discs, the authors deduce that the half-opening
angle of the disc is consistent with zero within the measurement errors, and does not
exceed 10◦ .
Levin and Beloborodov (2003) propose that a recent burst of star formation has
occurred in a dense gaseous disc around Sgr A*. In such a self-gravitating disc the parti-
cle density is at least four orders of magnitude larger than that of the densest molecular
clouds in the Galaxy, enough to overcome the tidal forces to start fragmentation. The
initial characteristic mass (Jeans mass) is accordingly high, about 1.8 M⊙ . Therefore, star
formation in a disc could naturally explain the observed dominance of high-mass stars.
Taking advantage of the adaptive optics at the VLT, Genzel et al. (2003a) could signifi-
cantly extend the sample of stars with proper motions. Following Levin and Beloborodov
(2003), Genzel et al. (2003a) find a second disc of stars (Figure 5.7). The two discs are
1.05
HeI 2.1127
293 (±35) km/s
1.3(±0.3) Å
1.00
Sν ν (Normalized)
HI Brγ 2.1661
464 (±20) km/s
0.95 5.8(±0.4) Å
-2
0.90
<1Å
0.85 FWHM=
1200(±80) km/s
EW=11.7(±2) Å
2.10 2.15 2.20 2.25
Wavelength (µm)
Figure 5.8. The ‘S-star’ cluster in the central arcsecond of the Galactic Centre. Left: integrated
SINFONI K-band (1.95–2.45 µm datacube with spectral classification of the S-stars: solid cir-
cles indicate early type stars, dashed circles indicate late-type stars). Right: average spectrum
of the bright (mK < 15, top) and faint (mK = 15 − 16, bottom) early-type stars (courtesy of
Eisenhauer et al., 2005).
inclined at large angles and counter-rotate with respect to each other. The authors also
find that the proper motions of the many more stars without radial velocities are not
randomly orientated but group at maximum absolute projected angular moment along
the line of sight:
Jz xvy − yvx
=. (5.17)
Jz,max (x2 + y 2 )(vx2 + vy2 )
where x, y and vx , vy are the sky coordinates and velocity components. Jz /Jz ,m ax =
+1, 0, −1 for clockwise tangential, radial, counter-clockwise tangential orbits. The group-
ing of the stars in absolute projected angular moment fits the two discs well.
Based on the spectra from integral field spectroscopy (see above), the stellar content
of the two discs is essentially the same, indicating that they formed at the same time.
Although other scenarios cannot yet be fully excluded, the ‘paradox of youth’ on a parsec
scale may be solved: the authors conclude that of the possible formation scenarios for
these massive stars the most probable one is that 5–8 Myr ago two clouds fell into the
centre, collided, were shock-compressed and then formed two rotating (accretion) discs,
which then fragmented and formed the stars.
5.3.5 SINFONI in the Galactic Centre: young stars and infrared flares
This paper by Eisenhauer et al. (2005) reports on 75 mas resolution, NIR imaging spec-
troscopy within the central arcsecond of the Galactic Centre. The authors find that to a
limiting magnitude of K ≈ 16, 9 out of 10 stars in the central 0.4′′ , and 13 out of 17 stars
out to 0.7′′ from the central black hole have spectral properties of B0–B9 main sequence
stars (Figure 5.8). The spectral classification is based on the detection of the Brγ and
He I lines. Different from the emission line stars in the stellar discs at 1–10′′ , the S-stars
show Brγ in absorption, which is characteristic of B-type stars. Another indication is the
lack of the typical CO band-head absorption of late-type stars (see Figure 5.3). Averag-
ing the spectra of the brighter and fainter stars, the signal-to-noise ratio is large enough
even to distinguish the spectral sub-type from the equivalent width (EW) of the He I
line. The bright stars (mK < 15) show strong (EW > 1 Å) absorption, typical for B0–B3
stars, while the fainter (mK = 15–16) stars have very little H eI absorption (EW < 1 Å),
typical of B5–B9 stars. The low K-band luminosity and the detailed shape of the Brγ
line indicate that the S-stars are main sequence and not giants or supergiants.
The fact that nine out of 10 S-stars are early-type (= young) stars makes the ‘paradox
of youth’ even stronger. A possible explanation could be stellar mergers. Such mergers
would probably spin up the stars. However, based on the He I line width, all brighter
S-stars have normal rotation velocities of about 150 km s−1 , similar to solar neighbour-
hood stars. Therefore, the authors rate the stellar merger scenario as less likely.
As for the emission line stars at 1–10′′ , the explanation for the ‘paradox of youth’ could
come from the dynamical properties of these stars. Combining the radial velocities from
the integral field spectroscopy with imaging data, the authors could derive improved
three-dimensional stellar orbits for six of these S-stars. The orientations of the orbits
appear random and the orbital planes are not co-aligned with those of the two discs of
massive young stars 1–10′′ from Sgr A*. The authors thus exclude the hypothesis that
the S-stars as a group inhabit the inner regions of these discs.
But have the S-stars ever been in a disc? One mechanism to change the direction of the
angular momentum is Lens–Thirring precession from the rotating black hole. The rotating
space time breaks the spherical symmetry, which would normally preserve the angular
momentum. The period of this Lens–Thirring precession is (e.g. Levin and Beloborodov,
2003):
) *2 ) *3/2 ) *3
6 3 × 106 M⊙ 1 − e2 r0
PL T = 5.8 × 10 a −1
yr, (5.18)
M 1 − 0.872 4.6 × 10−3 pc
where a is the dimensionless spin parameter (a = 1 for maximum spin), M the mass of
the central black hole, and r0 and e the semi-major axis and eccentricity of the orbit.
For a spin parameter of a = 0.52 (Genzel et al., 2003b), however, the lifetime (10 Myr)
of two out of the six B-stars is too short for Lens–Thirring precession to account for the
change in angular momentum. Therefore, other mechanisms are necessary to explain the
S-stars (see also the excursion on the ‘paradox of youth’).
One potential explanation is a dissolving star cluster with an intermediate mass black
hole (IMBH), possibly formed by runaway merger in the cluster core. But it is not clear
if an IMBH can have such a tightly bound cluster, and the IMBH would have to be two
orders of magnitude more massive than can be explained by formation from a runaway
merger. Another explanation might be a massive binary exchange, where the S-stars
were originally in a massive (≈ 100 M⊙ ) binary system and switched partner in a close
interaction with the black hole. However, there is no evidence of such massive WR/O stars
on highly eccentric orbits, and more massive stars should also have been captured, which
is not observed. The last scenario discussed here is exchange capture with stellar black
holes, in which the B-stars play billiards with equal mass stellar black holes in a central
cusp. The advantage of this scenario is that it would favour the observed eccentric orbits
and would prefer B-stars because they best match the mass of the about 10 M⊙ stellar
black holes. This scenario is also supported by X-ray observations indicating the presence
of about 104 neutron stars and stellar black holes in the Galactic Centre region. However,
this scenario requires a large reservoir of B-stars at larger radii, which are not found.
In summary, the ‘paradox of youth’ for the S-stars has not yet been solved, although the
observations indicate that these stars are probably brought into the central light-month
by strong individual scattering events.
a)
b)
d)
c)
Figure 5.9. The four stages of star formation: a) cores form within molecular clouds; b) pro-
tostar with surrounding nebular discs; c) stellar wind creating bipolar outflow; d) newly formed
star with circumstellar disc (courtesy of Shu et al., 1987).
5.4 Star formation

This section is about integral field spectroscopic studies related to star formation. A
recent review of this topic can be found in Garcı́a (2006). Compared to Galactic Centre
observations, the impact of integral field spectroscopy on star formation research has been
much smaller. A possible reason might be the wide variety of objects and questions so that
the early integral field spectroscopic studies are less coherent and less straightforward.
Therefore, the papers presented in this section are not ordered by publication date but
are grouped by topic. Some papers (for example on planetary nebulae) have been omitted
to present the subject in a more coherent way. After a brief introduction to theory of
star formation, this section discusses integral field spectroscopic observations of T Tauri
stars and their immediate vicinity, their jets and the emerging Herbig–Haro objects. This
is followed by a discussion of studies related to the formation of high mass stars, and
finally on the interaction between the winds and radiation of high mass stars with nearby
protoplanetary discs.
Excursion: star formation in a nutshell

This section introduces the basic concepts of star formation theory. Many good lectures
are available on the internet for download, for example from Cornelis Dullemond.1 There
are also a number of good review articles, for example Shu et al. (1987). Specifically
recommended for further reading are the conference proceedings ‘Protostars and Planets
I–IV’ published by The University of Arizona Press. The different phases of star formation
are illustrated in Figures 5.9 and 5.10.
The first phase is the collapse of a molecular cloud (Figure 5.9a). Collapse occurs if
the mass and density are large enough to overcome the thermal and magnetic support
of the cloud. The so-called Jeans mass, MJ , defines the limit for a thermally supported
cloud:
3 1
MJ ∝ T 2 R − 2 . (5.19)
1
http://www.mpia-hd.mpg.de/homes/dullemon/lectures/starplanet/
Figure 5.10. The formation of discs and outflows: a,b) gravitational collapse of a magnetically
supported molecular cloud; c) formation of an accretion disc; d) magnetically threaded disc;
e) bipolar outflow; f) head of the jet (courtesy of Cornelis Dullemond).
The critical mass for magnetic support is:

MB ∝ BR2 , (5.20)
where T is the temperature of the cloud, R the radius and B the magnetic field. The
major barrier to continued collapse is that the magnetic field is frozen in the molecular
cloud, even if the fraction of ionized gas is very small (Figures 5.10a and 5.10b). The
contraction also increases the magnetic field, and the collapse would stop as soon as the
critical mass, MB , is reached. This barrier is overcome by ambipolar diffusion, the slow
decoupling of the neutral molecules and dust particles from the electrons and ions frozen
in the magnetic field.
In the next phase, a circumstellar disc is formed. As the cloud rotates, infalling matter
collides with matter from the other side, forms shocks and converts kinetic energy into
heat (Figure 5.10c). This heat is radiated away, so that the matter cools and settles on
the mid-plane.
Because of friction, the matter in the disc slowly loses its angular momentum and moves
inwards. The plasma in the disc pulls the magnetic field with it, forming a magnetically
threaded disc (Figure 5.10d). The magnetic field creates an effective gravitational poten-
tial along the field line, and as soon as the angle between the field lines and the disc
gets below 60◦ , matter will start to outflow along the magnetic field, like in a slingshot.
These are the so-called ‘disc winds’. Further in – but still at a distance significantly larger
than the stellar radius – the interaction of accreting mass with the magnetosphere of the
protostar truncates the gas disc. At that radius, Rx , the disc accretion divides into a
funnel inflow on to the star, and an outflow called the ‘x-wind’.
As the gas wants to preserve its angular momentum, it will bend and wind the field
lines, thus confining the outflow. This bipolar outflow (Figure 5.10e) is surrounded by a
bubble of hot gas. A terminal shock forms at the head of the jet (Figure 5.10f). Most
of the energy is dissipated in a stand-off shock at the boundary between the jet and
the external medium. This external matter is pushed away and forms a bow shock. The
shocks in the ambient medium are observed as Herbig–Haro objects at distances up to
thousands of astronomical units. The jet is flanked by a region with turbulent mixing
between old jet material and the entrained environment. As the jet material is much more
tenuous than the external medium, the propagation of the bow shock is much slower than
the jet.
The protostar in the centre is a self-gravitating ball of hot gas continuously being
fed by accreting matter. Initially, there is no nuclear energy production, but the star is
heated from the gravitational energy released in the accretion and contraction (Figure
5.9b). The heating is countered by cooling from radiation, and the star contracts on the
Kelvin–Helmholz timescale until the central density and temperature are high enough to
>
start deuterium burning at T ∼ 106 K. At this stage, the star enters it pre-main sequence
phase (Figure 5.9d). The star becomes fully convective, and inflates again up to 3–5 R⊙ .
In a second, slow contraction the central temperature and density then increase until
hydrogen burning starts, and the star becomes a main sequence star.
Excursion: T Tauri and RW Auriga stars

The so-called T Tauri and RW Auriga stars are examples of low mass pre-main sequence
(PMS) stars. They are named after their prototypes, and are classified from the follow-
ing observational characteristics. The brightness is irregularly variable, typically with
variations of 1 mag within days. The stars also show strong emission lines from ion-
ized H, Ca and other elements. Their spectral types are in the range F–K (indicating
low-mass stars), and they are located above the main sequence in the Hertzsprung–
Russell diagram. These stars are found in star-forming regions, indicative of their young
age.
A major problem for the interpretation of the characteristic signatures is that these
signatures arise from multiple regions, including the stellar photosphere, circumstellar
discs, jets and shocks. In addition, many of these stars live in multiple systems so that
it is difficult to disentangle the components. This is particularly true for the ‘prototype’,
T Tau.
5.4.1 A near-infrared spectral imaging study of T Tau

This paper by Herbst et al. (1996) presents K (2.01–2.42 µm) and H (1.5–1.8 µm) band
imaging spectroscopy of the complex environment of the young stellar object T Tau.
T Tau is a pre-main-sequence binary system, with an optically visible primary star, and
a highly obscured infrared component, separated by approximately 0.6′′ (see Figure 5.12).
The observations by Herbst et al. (1996) reveal at least five distinct sources of molecular
hydrogen emission in the central 4′′ × 4′′ of the system: diffuse knots on primary (N)
and infrared (S) components, a bright NW knot (T Tau NW) with a SE counterpart,
and an east–west jet. The ratio of various H2 lines (see excursion on H2 line diagnostics)
indicates that all components are shock-heated.
The strongest source in molecular hydrogen emission is T Tau NW. It is detected in 11
different lines. The spectrum is typical for a Herbig–Haro Object (Figure 5.11a). The H2
spectral-line ratios point to shock heating with temperatures in the range 1000–2500 K
(Figure 5.11b). However, T Tau NW is also detected in two H-band forbidden transitions
of [Fe II] (square brackets indicate forbidden lines). Normally, a shock that produces
[Fe II] quickly destroys H2 . This implies that we see a range of excitation conditions. The
authors also find a slight offset between H2 and [Fe II], indicating that [Fe II] is coming
from the tip of the shock where the energy is dissipated, and that the H2 emission is
flanking the shock. An apparent SE counterpart supports the Herbig–Haro hypothesis.
The authors further detect an H2 filament towards the west, which appears to be a
signature of a second outflow and its counter-jet to the east. The near perpendicularity
a) 1.4
1.3
T Tau NW K
1.2
1.1
1.0
0.9
0.8
0.7
2 2.1 2.2 2.3 2.4
1
b) c) 6 v=0 v=1
0 TTau NW 4
1-0 S(0)
5
1-0 S(1)
3
1-0 S(2)
1-0 S(7)
2
1
2-1 S(3)
2-1 S(2)
4 0
7
3
T = 2500
6
H2
2-1 S(1)
T = 2000 2 5
T = 1500
1 4
3
2
0 1
0 2000 4000 6000 8000 10000 0
Figure 5.11. Emission from molecular hydrogen in T Tau: a) K-band spectrum of T Tau NW.
The flux ratio of H2 ν = 1 − 0 S(1) and H2 ν = 2 − 1 S(1) (highlighted) give a simple first estimate
for the excitation mechanism; b) excitation diagram (both figures courtesy of Herbst et al., 1996);
c) level diagram with H2 ν = 1 − 0 S(1) transition.
of T Tau NW and the west filament suggests that one star of the binary system produces
each outflow. Indirect evidence suggests that T Tau NW is associated with the primary
star, and the west filament with the infrared companion. The thermal, diffuse H2 emission
observed at the location of the stars could be caused either by collimation of the NW
and W outflows or by accretion shocks, which would also explain the large extinction
and infrared luminosity of the companion.
Excursion: molecular hydrogen lines in the infrared

The vibrational and rotational excitations of a molecule determine its spectrum in the
infrared. The electronic excitations have much higher energies and are only observed in
the visible or at even shorter wavelengths. The energy levels for vibrational and rotational
excitations are:
) *
1
Ev ib = hν0 ν + (5.21)
2
h2
Er ot = J (J + 1) , (5.22)
8π 2 I
where ν and J are the vibrational and rotational quantum numbers, respectively, h the
Planck constant, ν0 the vibrational frequency, and I the moment of inertia. Figure 5.11c
illustrates the energy levels for the hydrogen molecule.
The line emission observed in the infrared results from vibrational transitions with
or without rotational changes. The nomenclature for the spectral lines gives the initial
and final vibrational quantum number, ν, the change in the rotation quantum number –
corresponding to a spectral branch – and the final rotation state, H2 ν = f rom − to
branch (final J). The change in rotation – the branch – is coded by capital letters: for the
Q-branch we have J → J, for the S-branch we have J → J − 2. Because H2 is lacking
a dipole moment, the R-branch (J → J − 1) is forbidden in this molecule.
Excursion: H2 line diagnostics

The Boltzmann equation gives the fractional number of particles per degenerate sub-level
ην,J occupying a set of ro–vibrational states ν, J for thermal equilibrium:
N (ν, J) Ntot −ϵ ν,J
ην,J ≡ = e (5.23)
gν,J Z(T ) T
where N (ν, J) and Ntot are the sub-level and total number of particles, respectively, gν,J
the degeneracy, and ϵν,J the energy of the state expressed in Kelvin.
Forming the ratio of ην,J with that of a reference state (ην 0 ,J 0 , ϵν 0 ,J 0 ) eliminates several
of these quantities, giving
ην,J 1
log = − (ϵν,J − ϵν 0 ,J 0 ) . (5.24)
ην 0 ,J 0 T
The corresponding log–normal plot is called an excitation diagram (see Figure 5.11b),
and the slope in this diagram is inversely proportional to the temperature. It can be used
to distinguish the two major excitation mechanisms: UV radiation with high excitation
temperature (> 2500 K), and shock heating with comparable low excitation temperature
(< 2500 K). A simple first estimate for the excitation mechanism can be derived from
the flux ratio of H2 ν = 1 − 0 S(1) – usually the strongest H2 line in the K-band –
and its H2 ν = 2 − 1 S(1) counterpart for the next higher vibrational transition: a flux
2 ν =1−0S (1)
ratio H
H 2 ν =2−1S (1) > 5 indicates shock heating, a flux ratio ≈2 indicates ultraviolet (UV)
excitation.
5.4.2 Spatially resolved imaging spectroscopy of T Tauri

The small separation of the stars in the T Tau system has prevented spectroscopy of
its individual components in the past. This paper by Kasper et al. (2002) presents near-
infrared data (H- and K-band) on this object taken at the Calar Alto 3.5 m telescope
using the ALFA adaptive optics system in combination with the MPE 3D integral field
spectrograph. The diffraction-limited observations allow us for the first time to clearly
resolve the spectra of the primary (N) star and the infrared (S) companion (Figure 5.12).
The spectrum of the primary T Tau N (Figure 5.12 top right) shows photospheric
absorption lines from Mg, Al and Ca. The measured line ratios are consistent with spec-
tral type K. However, the equivalent width of the lines is significantly smaller, requiring
a large amount of continuum veiling, for example from the hot inner part of a circum
stellar disc. This K band excess is of the order 80%.
The spectrum of the infrared component T Tau S (Figure 5.12 bottom right) is very red
and does not show photospheric lines. The stellar photosphere must be extremely veiled,
if not completely invisible. The Bracket line emission from T Tau S and T Tau N could
not be resolved with the 3.5 m telescope, implying that it arises within 6 AU from the
stars. The line ratios between Brγ (2.166 µm) and Br10 (1.736 µm) towards T Tau N and
T Tau S are similar, suggesting comparable selective extinctions of AV ≈ 6 toward the
2.05 2.15 2.25 2.35
(telluric) H2 Brγ 3.5
T Tau N
3.3
T Tau N 3.1
2.9
2.7
0.69'' 1.45
1.35
12CO 1.25
T Tau S
Br10
T Tau S 1.15
1.05
Mg Na Ca Mg
0.95
2.05 2.15 2.25 2.35
Wavelength [µm]
Figure 5.12. Integral field spectroscopy of T Tau. Left: Br10 (1.736 µm) image; right: K-band
spectrum of T Tau N (top) and T Tau S (bottom) (courtesy of Kasper et al., 2002).
respective emitting regions. All this indicates that T Tau S is not stellar light reddened by
strong absorption from cold dust, but that the emission arises from hot dust. Therefore
a protostar with an infalling envelope seems an unlikely explanation for T Tau S. On the
other hand, the results are consistent with a model that describes T Tau S as a pre-main
sequence star surrounded by a small edge-on disc. The disc is observed at an inclination
such that it obscures the star, at the same time leaving the line of sight towards the
Bracket emission from the polar regions relatively unobscured.
Excursion: microjets in pre-main sequence stars

Outflows and jets are common phenomena in pre-main sequence stars. In particular, we
find so-called microjets, which are only a few hundred AU long, and which are launched
at 0.1–10 AU radius. The microjets show forbidden line emission and high velocities.
! What is driving the jets? There are a number of competing explanations (see excur-
There are two long-standing questions on microjets:
sion on star formation): disc winds, x-winds and stellar winds. Unfortunately, the
available magnetohydrodynamical models do not predict line ratios, and simple spec-
troscopy cannot answer the question. Instead, the models predict velocities, densities
and magnetic field strengths. We thus have to measure quantities like diameter, den-
sity, terminal velocity and rotation speed to identify the driving mechanism of the
! What is the heating mechanism? Given the small launching radius (0.1–10 AU) of the
jets.
jets, dramatic adiabatic cooling should occur in the expansion region, and extended
forbidden line emission should not be detectable. This is in disagreement with the
observations. Several heating processes are discussed for the microjets: shocks against
the environment or within the jet, turbulence in a viscous mixing layer, ambipolar
diffusion heating and compression by jet instabilities.
Several integral field spectroscopic studies have been aiming at resolving these questions.
Disk Wind (Macc = 10−5 Msun yr−1) convolved with 14 AU beam
III= 0.7''–0.9'' II= 0.5''–0.7'' I = 0.3''–0.5''

Vshift(km/s)
Garcia et al (2001) 0
C
100
−20
Jet FWHM (AU) in SU [6731]

−40
DG Tau
Vshift(km/s)
B −0
50 CW Tau −20
HH 30
HL Tau −40
RW Aur
A
Vshift(km/s)
0
−20
0
0 50 100 150 200
−40
Projected distance from source (AU)
80 60 40 20
Distance R across the jet (AU)
400 [O I] RW Aur
Projected Vrad (km/s)
200
0
−200
−400 Sol A α = 0.003
−200 −100 0 100 200
Projected distance along jet (AU)
Figure 5.13. Comparison of the RW Aur microjet with cold disc wind models. Top left: jet full-
width-at-half-maximum (FWHM) measured from the [SII]λ6731 line; right: transverse velocity
shifts between symmetric positions on either side of the jet axis (both figures courtesy of Douga-
dos et al., 2003); bottom: position–velocity diagram measuring the terminal velocity (courtesy
of Garcı́a, 2006).
5.4.3 The RW Aur microjet: testing MHD disc wind models

This paper by Dougados et al. (2003) aims at answering the question of what is driving the
jets (see excursion). Based on observations with the adaptive optics bonnette PUE’O and
the OASIS (McDermid et al., 2004) spectro-imager at the CFHT, the authors conduct a
detailed comparison of the RW Aur microjet on scales < 200 AU with predictions from
cold disc wind models.
The observed collimation and width (≈ 30 AU) of the microjet of RW Aur can well be
modelled with cold disc winds with moderate to high ejection efficiencies (Figure 5.13
top left). However, the observed rotation signature – transverse velocity shifts between
symmetric positions on either side of the jet axis – of < 10 km s−1 is significantly below
what is expected from disc wind models (Figure 5.13 right): cold disc winds, extending
radially out to > 1 AU, would have transverse shifts of 30–40 km s−1 . Cold disc wind
models are also in conflict with the observed terminal velocities of the RW Aur jet
(Figure 5.13 bottom left). While the observed terminal velocities are ≈ 150–350 km s−1
(Garcı́a, 2006), the disc wind models predict much higher velocities of ≈ 500 km s−1 . The
x-wind models, on the other hand, would predict velocities that were too low.
So the answer to the question of what is driving the jets is still pending. There is
evidence in favour of and against all considered mechanisms, and more observations and
modelling are required to close the topic.
5.4.4 Sub-arcsecond morphology and kinematics of the DG Tauri jet in the

[OI]λ6300 line and DG Tau: a shocking jet?
These two papers by Lavalley et al. (1997) and Lavalley-Fouquet et al. (2000) present
the first integral field observations of the DG Tauri jet. The goal is to answer the second
question of what is heating the jet (see excursion).
Position ('')
2
6
0 2 4 0 500
R.A. offset (arcsec) Velocity (km/s)
[OI] [Ha] [N II] [SII]
2 HV
0
Offset (arcsec)
2 IV
2 LV
0 2 4 0 2 4 0 2 4 0 2 4
Offset (arcsec)
Figure 5.14. DG Tauri jet. Top left: deconvolved [O I]λ6300 map with an effective resolution
of 0.35′′ ; top right: position–velocity diagram along the jet axis, symbols indicating the centroid
velocities of the multi-component line fit (both figures courtesy of Lavalley et al., 1997); bottom:
channel maps for [O I]λ6300, Hα, [N II]λ6583 and [S II]λλ6716,6731 lines at [−400, −250] km s−1
(top), [−250, −100] km s−1 (middle), [−100, +10] km s−1 (bottom) (courtesy of Lavalley-Fouquet
et al., 2000).
The observations were obtained with the Tiger and OASIS instruments at the CFHT
in 1994 and 1998 and cover the [O I]λ6300, Hα, [N II]λ6583 and [S II]λλ6716,6731 lines
(the numbers following λ indicate the wavelength in Å). The morphological structure
of the mass outflow is revealed with unprecedented spatial resolution of approximately
0.35–0.5′′ . The deconvolved [O I] map (Figure 5.14 top left) shows an unresolved inner
peak containing two thirds of the total line flux, followed by a collimated jet-like body
extending out to ≈ 1.5′′ from the star, two resolved knots at distances of 2.7′′ and 4′′ ,
and a counter-jet. The unresolved inner peak is slightly displaced (≈ 20 AU) from the
star. It is apparently stationary and possibly linked to the initial jet collimation. The
size limit and absolute photometry for this peak allows Lavalley et al. (1997) to set
new constraints on the jet mass loss rate of 1.3 × 10−8 – 6.5 × 10−6 M⊙ yr−1 , lowering
previous estimates by a factor 5 to 25. From the [O I], Hα, [N II] and [S II] channel maps at
different velocities (Figure 5.14 bottom), the fast jet core appears surrounded by a slower-
moving flow or cavity. The two velocity components are analysed in more detail from the
position–velocity diagram (Figure 5.14 top right). A multiple component line fit for each
position clearly shows the two distinct velocity components at distances less than 1.5′′ :
the high-velocity component quickly accelerates up to 350 km s−1 ; the slow component
follows the fast component at proportionally slower velocities up to about 100 km s−1 .
Accordingly, the slow component is interpreted as circumstellar material entrained in the
jet. Beyond 1.5′′ (200 AU) only one velocity component is detected, and the jet starts
wiggling around (Figure 5.14 top left and bottom). The knot at 2.6′′ shows a curved
morphology (Figure 5.14 top left). This and the transverse velocity gradient are strongly
suggestive of a resolved bow shock. Radial velocity jumps of ≈ 50 km s−1 are observed
between the bright knots at 1′′ and 3.6′′ . In order to determine the heating process of
the jet, Lavalley-Fouquet et al. (2000) also compare the measured [N II]λ6583/[O I]λ6300
and [S II]λ6716/[S II]λ6731 versus [S II]λ6731/[O I]λ6300 line ratios with available model
predictions for shocks, ambipolar diffusion and turbulent mixing: the measurements are
best reproduced by shocks with speeds of ≈ 50–100 km s−1 and pre-shock densities of
≈ 105 –103 cm−3 .
Combining the evidence – the bow shape structure of the knot, the velocity jumps at
1–4′′ , the wiggling beyond 1.5′′ , the line ratios and the variability in the ejection velocity
reported elsewhere – indicates that ‘shocks’ is the answer to the question of what is
heating the microjets.
5.4.5 The three-dimensional structure of HH 32 from GMOS IFU spectroscopy

In this paper Beck et al. (2004) present high-resolution spectroscopic observations of the
Herbig–Haro object HH 32 made with the GMOS IFU (Allington-Smith et al., 2002) at
Gemini North Observatory.
HH 32 is one of the brightest sources in the original catalogue of Herbig and Haro.
Previous studies have shown that HH 32 has a high excitation spectrum with strong
emission from H, He and a wide range of metals. It also exhibits H2 emission from shocks
as the flow encounters ambient cloud material. The outflow is clearly associated with a
T Tauri star. Several condensations up to several 10′′ have been detected. The three-
dimensional spectral data from Beck et al. (2004) cover a field of ≈ 9′′ × 6′′ centred on
the HH 32 A knot complex. The spectral range is 4820–7040 Å. The position-dependent
line profiles and radial velocity channel maps of the Hα line show a variety of spatial
and spectral structures. The most prominent features are double-peaked radial velocity
profiles, two arclike structures at low and high velocities and two compact components
at intermediate velocity. These qualitative findings already indicate that HH 32 A is a
combination of at least two bow shocks.
The authors also present various line ratio velocity channel maps from the [O III]λ5007,
Hα, [O I]λ6300, [N II]λ6583 and [S II]λλ6716,6731 lines. The line emission and the line
ratios vary significantly on spatial scales of ≈ 1′′ and over velocities of ≈ 50 km s−1 . At
low velocities, the density tracer (see excursion on line ratios) [S II]λ6716/[S II]λ6731 is
diffuse and extended. The ratio of ≈ 0.7–1.0 corresponds to an electron density of ≈ 230–
150 cm−3 . The line ratio is getting more compact and is dropping at higher velocities.
At velocities above ≈ 110 km s−1 , the observed line ratio in the compact component is
≈ 0.55, corresponding to an electron density of ≈ 650 cm−3 . The line ratio [N II]/Hα –
which traces the excitation temperature (see excursion) – peaks in the diffuse, low-density
regions. We thus see high-density knots surrounded by a highly excited, low-density
plasma. As expected from a shock scenario, the line ratio [O III]/Hα – which traces the
shock velocity – increases with the observed velocity in the knot.
To interpret the observations, Beck et al. (2004) compare the measured channel maps
and line ratios with a ‘3/2-dimensional’ bow shock model (see excursion). The model is
qualitatively successful at reproducing the general features of the radial velocity channel
maps, for example the transition between a single arclike feature at low radial velocities
and a more concentrated and centrally peaked condensation at high radial velocities. But
0 200 400 600 800 1000 0 0.1 0.2 0.3 0.4 0.5
−20. −5. +10. +25. −20. −5. +10. +25.
+41. +56.. +72. +87. +41. +56. +72. +87.
+103. +118. +134. +149. +103. +118. +134. +149.
+165. +180. +196. +211. +165. +180. +196. +211.
+227. +242. +258. +273. +227. +242. +258. +273.
+289. +304. +320. +335. +289. +304. +320. +335.
+351. +364. +382. +397. +351. +366. +382. +397.
1.0"
Figure 5.15. Herbig–Haro object HH 32 A. Left: observed Hα radial velocity channel maps;
right: channel maps computed from a 3/2-dimensional bow shock model (courtesy of Beck et al.,
2004).
r
z
zmax vflow
vf low
vshock
Figure 5.16. Geometry of a 3/2-dimensional shock model.
the model does not show the same complexity as the data, and in particular it fails to
reproduce the line ratios. The reason for the failure is that the parameterization is much
too simple for the spatially complex region. Another shortcoming of the 3/2-dimensional
model is that it treats the shock as locally planar, i.e. by assuming that post-bow shock
cooling distance is small compared to the bow shock. In HH 32 A, however, the cooling
distance is comparable to the size of knots.
In summary, the integral field spectroscopic observations of HH 32 A reveal two or
three bow shocks along the redshifted body of the HH 32 outflow.
Excursion: 3/2-dimensional bow shock model

The so-called ‘3/2-dimensional’ models simplify the complex three-dimensional structure
! a simple axially symmetric surface, z ∝ rp ;

of a bow shock by assuming:
! that the local shock velocity v

sho ck is the projected pre-shock flow velocity vflow
! a finite shock size described by z

perpendicular to the shock surface;
! that the shock is locally described by a planar shock.

m ax ;
Figure 5.16 illustrates the geometry of the 3/2-dimensional model. The result from such
a 3/2-dimensional model for the Herbig–Haro object HH 32 A is shown in Figure 5.15.
Excursion: line ratio diagnostics in the optical
The optical wavelength range is rich in atomic lines caused by transitions between elec-
tronically excited states. The strength of an emission line is sensitive to the physical
conditions in the emitting region. Therefore line ratios provide a diagnostic tool for
deriving temperature, densities and other properties. A good introduction to this topic
can be found in Osterbrock (1989). Here are a few examples of diagnostic line ratios used
! [O I I I ]λλ(4959+ 5007) and [N I I ]λλ(6548+ 6583) : [O III] and [N II] are amongst the best
in the analysis of stellar jets and Herbig–Haro objects:
[O I I I ]λ4363 [N I I ]λ5754
examples of ions that have emission lines from two different upper levels with signif-
icant different excitation energies and for which both lines are easily observable in a
single spectrum. For low densities < 105 cm−3 , where collisional excitation is negligi-
ble, the level population follows the Boltzmann formula, and the line ratio traces the
excitation temperature (larger ratio indicating lower temperature). At high densities
> 105 cm−3 collisional de-excitation begins to play a role. Because the lower excita-
tion state has a considerably longer lifetime – a lower radiative transition probability
– it can be more quickly collisionally de-excited, thus weakening its line emission. In
! [O I I ]λ3729 and [S I I ]λ6716 : these line ratios are examples of density tracers. Because
this case the line ratio traces the density (higher ratio indicting lower density);
[O I I ]λ3726 [S I I ]λ6731
the two lines are emitted from levels with nearly the same excitation energy, the exci-
tation rate is almost independent of temperature. The relative strength of the opti-
cal transitions is set by the competing collisional de-excitation, which de-populates
the state more efficiently with the longer lifetime, i.e. a smaller radiative transition
probability (see above). Because this collisional de-excitation is strongly enhanced at
higher densities, the above line ratios are tracing directly the density of the plasma.
! [N I I ]λ6583 and [O I ]λ6300 : if the emission lines of the above indicators are too weak
A high line ratio indicates a low electron density;
Hα Hα
to be observed, line ratios from different species and ionization states can be used
to determine the physical parameters of the plasma; for example, the ratio of lines
with high excitation energies (e.g. [N II]λ6583 and [O I]λ6300) and the low excitation
Hα line trace the excitation temperature of the plasma. A high line ratio indicates
a high excitation temperature.
Excursion: high-mass star formation

>
Other than for low-mass stars, the formation of high-mass stars (∼ 20 M⊙ ) is poorly
understood and little explored observationally. The problem is that high-mass stars evolve
very quickly – on a Myr timescale – so that nuclear fusion starts before the end of
accretion. The observational problem is that at the time when the high-mass stars clear
their circumstellar clouds and get visible the formation process has already finished. The
theoretical problem is that the radiation from the star will stop its growth, and it is not
clear how the stars can acquire enough material before this happens. Several scenarios
have been proposed for the formation of high-mass stars: through disc accretion (similar
to low-mass stars), via collision of stars, and from Bondi accretion (collecting material
from travelling through the interstellar medium). A possible route to test the scenarios
is to search for accretion discs and associated outflows and jets.
5.4.6 Collimated molecular jets from high-mass young stars: IRAS 18151−1208
This paper by Davis et al. (2004) presents integral field spectroscopy of the high-mass
protostar IRAS 18151−1208, its jet and the associated shocks. IRAS 18151−1208 is a
class I young stellar object with a double-peaked spectral energy distribution dominated
0.3
0.2
Flux (Jy)
0.1
0
0.008 Brγ co 3–1
2–0 4–2
5–3
6–4
7–5
0.004
H2 0
1.6 1.8 2 2.2 2.4
Wavelength (µm)
0.008
1–0S(1)
Q–branch
0.007 Knot/Region B3
0.006
0.005
Flux (Jy)
0.004
1–0S(0)
1–0S(2)
2–1S(1)
1–0S(6)
1–0S(9)
1–0S(7)
3–2S(3)
0.003
2–1S(2)
1–0S(8)
2–1S(3)
0.002
0.001
0
1.7 1.8 1.9 2 2.1 2.2 2.3 2.4
Wavelength (microns)
Figure 5.17. IRAS 18151−1208. Top left: K-band continuum image of the IRS1 high mass
protostar; top middle: south-eastern structure pointing in the direction of the large scale jet;
top right: spectrum of IRS1; bottom: H2 ν1 − 0S(1) line maps and spectrum of a typical knot in
the jet (courtesy of Davis et al., 2004).
by cold (peak emission at 100 µm) and warm (peak emission at 20 µm) dust. Its bolo-
metric luminosity is ≈ 20,000 L⊙ , indicating that IRAS 18151–1208 is a high-mass star
in formation. Near-infrared images in the emission lines of molecular hydrogen H2 have
revealed two collimated jets in the IRAS 18151−1208 region, one of which is almost a
parsec in length. If these jets originate from the young stellar object, this would imply
that high-mass stars form via disc accretion. The K-band image also shows a bright cen-
tral peak (Figure 5.17 top left), denoted IRS1, roughly at the position of the mid-infrared
source and aligned with one of the jets. But is IRS1 really the high-mass star? And is
the jet really coming from that star? Unfortunately the error ellipse for the position
of the mid-infrared source is about 25′′ × 7′′ . The position of IRS1 in the near-infrared
colour-magnitude diagram supports the view that this star is indeed the high-mass star.
But this photometric identification is not conclusive, because emission from circumstel-
lar discs and dust extinction prevents the straightforward conversion from near-infrared
colours and magnitudes to temperature and luminosity.
The integral field spectroscopy from Davis et al. (2004) shows the H- and K-band
spectrum of IRS1 steeply rising at longer wavelengths (Figure 5.17 top right), indicating
a deeply embedded object. IRS1 also shows CO band-heads in emission, tracing dense
molecular gas (nH ≈ 1012 –1013 cm−3 and T ≈ 2000–5000 K) probably heated by UV pho-
tons impinging on the inner accretion disc surface. Alternatively, the CO emission could
come from the outer regions of accreting funnel flows, where the gas is heated by adiabatic
compression. The line maps from the [Fe II]λ1.644 µm and H2 emission (Figure 5.17 top
middle) show an ≈ 1′′ -long south-eastern structure extending from IRS1 in the direction
of one of the large-scale jets. Additional echelle spectroscopy supports the hypothesis that
this feature is indeed the base of the jet. The distant jet components show only H2 (Figure
2 ν =1−0S (1)
5.17 bottom right), but not [Fe II], indicating low excitation. The line ratio HH 2 ν =2−1S (1)
is approximately 7–11, typical for shock heating (see excursion on H2 line diagnostics).
The analysis of the excitation diagram gives a temperature of ≈ 2000–2600 K. The knots
have a bow-like morphology (Figure 5.17 bottom left), and the kinematics are compa-
rable to what is seen in Herbig–Haro objects. In summary, Davis et al. (2004) find that
the knots in the IRAS 18151−1208 jets are similar to their counterparts from low-mass
protostars. The close association between the H2 features and the high-velocity CO emis-
sion observed at radio wavelengths suggests that the CO represents gas entrained by the
collimated jets. From the mass and momentum of the molecular gas and the luminosity
of the H2 features, it is clear that the flow must be powered by a massive source. To all
intents and purposes, the molecular jet appears to be a scaled-up version of jets from
low-mass young stellar objects.
Collectively, the observations add further support to the idea that massive stars are
formed through vigorous disc accretion, and that massive protostars drive collimated jets
in their earliest stages of evolution.
5.4.7 GEMINI multi-object spectrograph integral field unit spectroscopy of the

167–317 (LV2) proplyd in Orion
This paper by Vasconcelos et al. (2005) presents high spatial resolution spectroscopic
observations of the proplyd 167–317 (LV2) in the Orion Nebula obtained with the GMOS
integral field unit at the Gemini South Observatory.
Proplyds are protoplanetary discs that are being exposed to an intense ultraviolet
radiation field which renders them visible. The characteristics of a proplyd are a bow-
shaped head that faces the ionizing star, a tail that is directed away from the source, often
a young star and a disc, sometimes seen in silhouette against a background H II region.
There are more than 100 proplyds in the Orion star-forming region, photoionized by
the Trapezium cluster. 167–317 (LV2) is a particularly bright proplyd. Previous studies
have found spectral signature of a collimated outflow, and its mass loss rate has been
estimated to approximately 8.2 × 10−7 M⊙ yr−1 .
Using integral field spectroscopy, Vasconcelos et al. (2005) have detected 38 forbidden
and permitted emission lines associated with the 167–317 (LV2) proplyd. The [N II]λ5755,
[N II]λ6548, Hα, [N II]λ6583, [He I]λ7065 and [Ar III]λ7135 are strong enough to construct
intensity and line profile maps. Figure 5.18 top left shows the intensity and line profile
map for Hα. In order to identify the jet of 167–317 (LV2) and to analyse its properties, the
authors extract velocity moment maps (see excursion). These velocity moments provide
an unbiased and robust simplification of the velocity substructure. Both Hα velocity
dispersion and skewness (Figure 5.18 top middle and right) peak towards the south-east,
indicating a redshifted wing in the profile. Indeed, the Hα line profile (Figure 5.18 bottom
left) shows a significant asymmetry, and subtracting a simple Gaussian fit reveals the red
and blue components (Figure 5.18 bottom right). The redshifted component is associated
with the previously reported jet. It has a receding velocity of about 80–120 km s−1 and
it is spatially distributed to the south-east of the proplyd. The approaching structure
may possibly be associated with a faint counter-jet with a velocity of −75 ± 15 km s−1 .
There is evidence that the redshifted jet has a variable velocity, with slow fluctuations as
a function of the distance from the proplyd. The line ratio [N I I[N
]λλ(6548+ 6583)
I I ]λ5754 in the jet is
6
approximately 10, corresponding to a particle density of ≈ 2 × 10 cm (see excursion on
−3
line ratio diagnostics in the optical). Vasconcelos et al. (2005) also provide an updated
mass loss rate for the 167–317 (LV2) proplyd from a simple model for the observed
Hα fluxes (see excursion): the mass loss for the photo-evaporated flow and the red-
shifted microjet are Ṁproplyd = (6.2 ± 0.6) × 10−7 M⊙ yr−1 and Ṁjet = (2.0 ± 0.7) ×
200 400 600 800
50 50
40 40
30 30
20 20
10 10
0 0
a) Hα
Flux (× 10−15 erg cm−2 s−1)
Flux (× 10−15 erg cm−2 s−1)

120 10
100 (6,7) 5
80
60 0
40 −5
20
0 −10
−200−100 0 100 200 300 −200 −100 0 100 200 300
Vrad (km s−1) Vrad (km s−1)
Figure 5.18. 167–317 (LV2) proplyd in Orion. Top: Hα intensity and line profile map (left),
flux weighted rms (middle) and skewness (right); bottom: typical Hα line profile (left); data
minus simple Gaussian fit (right), revealing a blue- and a redshifted component (courtesy of
Vasconcelos et al., 2005).
10−8 M⊙ yr−1 , respectively. For a typical proplyd mass of ≈ 0.1 M⊙ , the mass loss rate
implies that the lifetime is less than ≈ 105 yr.
Excursion: velocity moments

A common problem of integral field spectroscopy is to extract the ‘important’ information
from the vast amount of data. Line profile maps (Figure 5.18 top left), which show
the spectrum for every pixel, contain all the information, but are difficult to read and
interpret. One possible way to condense the information is to fit the line profile for every
pixel and to interpret the fit parameters. However, line fit has the disadvantage of being
biased from the assumed number of lines and their profile (for example, Gaussian or
Lorentzian), and is not always numerically stable (returning the wrong optimum). An
alternative is velocity (v) moments. The first three moments are the mean < v >, variance
∆v 2 and skewness ∆v 3 :
!
1
<v>= Iv vdv (5.25)
I
!
1
∆v 2 = Iv (v− < v >)2 dv (5.26)
I
!
1
∆v 3 = Iv (v− < v >)3 dv (5.27)
I
These quantities have the advantage of being unbiased and numerically stable. The vari-
ance indicates the width of the line and the skewness describes the asymmetry of the
profile (positive skewness indicates a redshifted wing). Figure 5.18 top illustrates the
technique for the velocity field of a protoplanetary disc. The result may be used as an
initial guess for subsequent line profile fitting.
Excursion: simple model for mass loss rate

The mass loss rate Ṁ from a proplyd can be estimated from a simple model that assumes
that the flow arises from a hemispherical wind that originates at the proplyd ionization
front with the sound speed, vS . Under these assumptions we can write:
Ṁ = 4πR2 ρ vS (5.28)
2
where R is the radius, 4πR the surface and ρ the density of the hemispherical ionized
region. The density, ρ, can be estimated from various line ratio diagnostics (see the rele-
vant excursion), or from the Hα luminosity LH α . The assumption for the latter estimate
is that the probability for the electron–proton recombination – a two-particle process –
is proportional to the square of the particle density, n:
ṅ = α n2 (5.29)
4 3
LH α = πR ṅ hν (5.30)
3
ρ = µ n mH , (5.31)
where µ is the mean molecular weight of the ionized gas, mH the mass of a hydrogen
atom, 43 πR3 the volume of the ionized region, ṅ the recombination rate (in particles per
cm3 and second), hν the energy of a Hα photon, and α ≈ 5.83 × 1014 cm−3 the effective
recombination constant for the Hα line. Combining the formulae we get the estimate for
the mass loss rate:
Ṁ = 4πR2 n µ mH vS (5.32)
5
3 LH α
n= (5.33)
4π α hν R3
Typical numbers for a proplyd are R ≈ 10 AU, n ≈ 106 cm−3 , and vS ≈ 10 km s−1 , imply-
ing a mass loss rate Ṁ ≈ 10−6 M⊙ yr−1 .
5.5 The Solar System

This section is about integral field spectroscopic studies of Solar System objects. Very
few publications have emerged from this field, probably because the Solar System is the
domain of space missions. Also, the scientific impact has been minimal. Nevertheless,
this section describes the early work on integral field spectroscopic observations of Mars,
because this study uses the very powerful principal component analysis to condense the
vast amount of information in the datacubes. The new adaptive optics-assisted instru-
ments are about to infuse new life into the field, and several papers are in the press or
in preparation. Therefore, the section also includes the discussion of a paper on Titan,
which does not fulfil the criterion of being published before the IAC Winter School in
November 2005.
5.5.1 Martian surface mineralogy from 0.8 to 1.05 µm TIGER spectra-imagery

measurements in Terra Sirenum and Tharsis Montes formation
This early paper by Martin et al. (1996) presents 0.8–1.05 µm spectra of Mars from
TIGER spectra-imaging. The primary goal of this investigation is to identify the physico-
chemical origins of the observed spectral variations, to map the surface based on the
1.30
a Mangala Valles 2 c
(0.34)
0.995 Daedalis Planum
0.825 1.20
Mangala Valles 1 (0.22)
(0.35)
(normalized to 1.0 at 0.85 µm)

0.86

Kest Mamnonis 0.93 0.985
1.20 (0.33) Terra Sirenum 2
Scaled reflectance
0.89 0.87
Scaled reflectance 0.98 (0.17)
0.905 North Uranius Patera
(0.28) 1.10 0.92
Sirenum Fossae
(0.21)
0.975
0.91 Amazonis 2
1.10 (0.28) Terra Sirenum 1
0.93
(0.25)
0.86 Amazonis 1
(0.30)
0.925 1.00
0.885 Thersis/West Arsia
(0.32) Terra Cimmeris
1.00 1.0 0.98
(0.19)
8500 9000 9500 10,000 10,500 11,000 8500 9000 9500 10,000 10,500 11,000
Wavelength (Ångstroms) Wavelength (Ångstroms)
1.1
Scaled reflectance
1.05
Endmember B (0.29)
1.0
Endmember A (0.13)
0.95
c
8500 9000 9500 10,000
Wavelength (Ångstroms)
Figure 5.19. Spectra of the Marian surface. Left: typical bright regions characterized by the
absorption band from ferric iron oxides at ≈ 0.86 µm; right: typical faint regions characterized
by the absorption band from proxine at ≈ 0.93 µm and ≈ 0.96 µm; bottom: end-members used
for the linear decomposition of the spectra (courtesy of Martin et al., 1996).
spectroscopic properties and to draw conclusions on the compositional heterogeneities of

bright and dark materials.
The measurements were obtained at Mauna Kea CFH Observatory during the 1990
opposition, and covered a systematic survey of the Tharsis volcanic region with a spatial
resolution of 200–250 km. To first order, the Martian surface can be characterized by low
and high albedo terrains. Figure 5.19 shows typical spectra of these regions. The spectrum
of the bright – high albedo – terrains is characterized by the absorption band of ferric iron
oxides (F e+3
2 O3 ) at 0.85–0.89 µm. Faint – low albedo – terrains show absorption from
pyroxine (XY Si2 O6 where X is a metal ion, and Y a smaller-sized ion) at 0.90–0.94 µm
(from ferrous pyroxine, X = Fe+2 ) and at 0.9–l.0 µm (from wollastonite, X = Ca+2 ). To
classify and characterize the ≈ 1500 spectra from the survey, the authors apply a prin-
cipal component analysis (see excursion), which decomposes the spectra into the ‘most
important’ components: the first principal component turns out to be indicative of the
slope in the spectrum, which can be interpreted as the relative absorption from ferric (at
shorter wavelengths) and ferrous (at longer wavelengths) minerals. The second principal
component indicates the bending of the spectrum and therefore measures the heterogene-
ity of the material. Next, the authors identify in the principal component diagram two
groups with typical spectra (called end-members), and look for the corresponding ter-
rains: the first spectrum is representative of a cratered rocky region with low reflectance,
2 2
PC
PC
2
1
α(X-ray), normalized
1
α(X-ray)
1.5 0
−1
1
−2
3 × 103 104 −2 −1 0 1 2
FWHM (Hb), km s−1 log FWHM(Hb), normalized
Figure 5.20. Geometric picture of principal component analysis (here illustrated for the X-ray
spectral index–Hβ line width relation for quasars). Left: measurements in natural units; right:
normalized units (mean subtracted, divided by standard deviation) and direction of principal
components (courtesy of Francis and Wills, 1999).
and – by comparing with laboratory spectra – corresponds to a mixture of ≈ 60% bulk

haematite (from the Greek for ‘bloodlike’ because of the red colour of its powder) and
≈ 40% of small (45–75 µm) particles of clinopyroxene (from Latin ‘ignis’ meaning ‘fire’,
common in igneous rocks from volcanoes). The second spectrum maps a highly fractured
terrain originating from lava flows covered with high albedo soils composed of active
dust deposits. This spectrum can be represented by a mixture of ≈ 90% haematite with
a large fraction of nanoparticles and ≈ 10% clinopyroxene. The two end-member spectra
are then used for the linear decomposition of the dataset. This mixture modelling reveals
that ≈ 90% of the observed terrains can be explained by a combination of clinopyroxene-
dominated and haematite-dominated material. In contrast, a linear mixture of low and
high albedo soils mentioned in previous Martian observations cannot explain the spec-
trum of the intermediate albedo soils. Both ferric and ferrous signatures are found in the
spectra of most dark and bright regions, demonstrating the strong heterogeneity of the
soils.
Excursion: principal component analysis

Principal component analysis (PCA) is a technique that can be used to simplify a dataset.
It is a linear transformation to a new coordinate system such that the greatest variance
in the dataset is along the first axis (‘first principal component’), the second greatest
variance along the second axis, and so on. In many cases the low-order components
contain the ‘most important’ aspects of the data and by showing which original variables
correlate with one another, principal component analysis often leads to new physical
insights. An introduction to the principal component analysis applied to astronomy can
be found in Francis and Wills (1999).
Geometric picture of principal components: The principal components can be inter-
preted as a series of linear least squares fits to a dataset, all mutually orthogonal: the
first principal component is an unconstrained minimum distance fit, the second principal
component a fit in a direction plane perpendicular to the first principal component, and
so on. Figure 5.20 illustrates the geometric picture of principal component analysis.
Algebraic picture of principal components: Here, we describe the data as a (N × P )
matrix X, where the element xi,j contains the ith measurement (i = 1, . . . , N ) of the
j th parameter (j = 1, . . . , P ), i.e. each row contains one spectrum, N is the number of
spectra and P the number of spectral channels:
⎛ ⎞
x1,1 x1,2 · · · x1,P →λ
⎜ x2,1 x2,2 · · · x2,P ⎟
⎜ ⎟
X = ⎜. .. .. .. ⎟ (5.34)
⎝ .. . . . ⎠ ↓
xN ,1 xN ,2 ··· xN ,P ID
Accordingly, the columns xi (i = 1, . . . , N ) contain the various measurements for a given

wavelength. The average value for a given wavelength is denoted x̄i . The principal compo-
nents of the dataset are then the eigenvectors of the (P × P ) covariance matrix Cov(X)
ordered by the size of the eigenvalues:
⎛ ⎞
cov(x1 , x1 ) · · · cov(x1 , xP )
⎜ cov(x2 , x1 ) · · · cov(x2 , xP ) ⎟
⎜ ⎟
Cov(X) = ⎜ . .. .. ⎟ (5.35)
⎝ .. . . ⎠
cov(xP , x1 ) ··· cov(xP , xP )
N
1 $/ 0/ 0
cov(xj , xk ) = xi,j − x¯j xi,k − x¯k (5.36)
N i=1
where j, k = 1, . . . , P . For our application, the principal components have the format of
a spectrum.
Excursion: seasons on Titan

Titan is the largest moon of Saturn. It has drawn public attention from the recent visit
and landing of the Cassini and Huygens mission in 2005. Titan’s atmosphere is strikingly
similar to the atmosphere of the pre-biotic Earth, dense in nitrogen with a pressure of
≈ 1.5 bar, and a surface temperature of ≈ 90 K. The second most abundant gas, methane
(≈ 5%), is close to its triple point, and takes part in a methane-based meteorological cycle
that is similar to the Earth’s hydrological cycle. A series of chemical reactions result in
the formation of aerosols in the upper stratosphere at ≈ 600 km. When the aerosols settle
into the troposphere (0–40 km), they serve as condensation nuclei for Titan’s methane
and ethane clouds. This is the starting point of an aerosol-climate feedback cycle: the
aerosols and clouds absorb and scatter the sunlight and therefore determine the thermal
structure of Titan’s atmosphere; in turn, the temperature gradients drive the atmospheric
circulation, which redistributes the aerosols and haze, closing the cycle.
5.5.2 Titan’s bright spots: multi-band spectroscopic measurement of

surface diversity and hazes
Ádámkovics et al. (2006) have obtained spatially resolved near-infrared spectra of Titan
that allowed them to derive the three-dimensional haze distribution in the atmosphere
and map the albedo of its surface.
The authors used the SINFONI integral field spectrometer with adaptive optics at
the VLT in February 2005 to obtain spatially resolved 1.45–2.45 µm spectra of Titan.
Figure 5.21 shows a sample of spectral images (top) and a typical K-band spectrum
(bottom left) extracted from the datacube. All of Titan’s observed near-infrared inten-
sity is sunlight scattered by aerosols or reflected from the surface. The dominant source
of opacity is the methane bands at wavelengths around 1.669 µm and 2.190 µm. There-
fore, the H- and K-band centres show the strongest absorption, and we observe sunlight
2.039 2.140 2.190
Tropospheric haze extinction enhancement

3.0
0.14
5K
2.5
0S, 68W, am = 1.17
asurf = 0.085 0.10
g = 0.68
I/F
2.0
0.06
1.5
0.02 K-band
1.0
2.00 2.05 2.10 2.15 2.20 2.25 −80 −60 −40 −20 0 20
Latitude
Wavelength [µm]
Figure 5.21. Integral field spectroscopy of Titan. Top: spectral images in the K-band probing
the surface (left), troposphere (middle) and stratosphere (right); bottom left: typical K-band
spectra (dashed line shows the best fitting radiative transfer model); bottom right: tropospheric
haze enrichment (courtesy of Ádámkovics et al., 2006).
scattered as aerosol haze high in the lower stratosphere (40–200 km). At shorter wave-
lengths (e.g. 1.588 µm and 2.039 µm) – towards the limits of the H- and K-bands – we
see only weak absorption, and the observations trace the albedo of the surface. Interme-
diate wavelengths (e.g. 1.611 µm and 2.140 µm) probe accordingly intermediate altitudes,
and the spectral images show the tropospheric (0–40 km) haze distribution. The strato-
spheric haze has a higher concentration near the north (winter) pole. In contrast, the
tropospheric haze is enhanced near the south pole (Figure 5.21 bottom right). Surface
albedos are recovered at 60 mas (375 km) resolution in both the 1.5 µm and the 2.0 µm
windows, nicely tracing the high-resolution surface maps from the Cassini spacecraft.
For the detailed analysis the authors apply a two-stream, plane–parallel, radiative trans-
fer model to fit the observed spectra (see excursion). This modelling provides a tool to
determine the vertical haze profile and retrieve latitudinal trends in both stratospheric
and tropospheric aerosol extinction: the stratospheric aerosol extinction is measured to
increase at a rate of ≈ 0.65 ± 0.05% per degree latitude from 40S to 60N. In contrast, the
tropospheric haze is confined near the south pole at latitudes above 40S.
Excursion: radiative transfer modelling

The analysis of the three-dimensional haze distribution in Titan’s atmosphere is based on
a radiative transfer model. In such a model, the software solves the two-stream radiative
transfer equation for multiple plane–parallel layers.
The model for the analysis of the Titan spectra is based on the atmospheric tempera-
ture and pressure profile measured in situ by the Huygens lander. The model takes into
account the scattering at the atmospheric haze, the absorption from methane, and the
reflection from the surface. A Henyey–Greenstein phase function is used to describe the
angular distribution of scattered light by the aerosol and is parameterized by the average
cosine of the scattering angle, g. As an intermediate result the model returns the so-called
contribution function, which is the fraction of photons at the top of the atmosphere that
arise from a particular altitude. After an additional limb correction, the model ultimately
returns the reflected spectrum.
The model fit has only two free parameters, the surface albedo and the characteristic
parameter, g, of the aerosol scattering phase function. To begin the fit, a series of spectra
is calculated by varying g and compared against the observed spectrum of the methane
absorption in the H- and K-band centres. Once g has been determined, the surface
albedos are varied to fit the spectra below 1.60 µm and 2.05 µm. Figure 5.21 (bottom
left) shows a typical K-band spectrum and the best-fitting radiative transfer model.
REFERENCES
Ádámkovics, M., de Pater, I., Hartung, M. ET AL. (2006), JGR, 111, 7
Bacon, R., Adam, G., Baranne, A. ET AL. (1995), A&AS, 113, 347
Bacon, R., Copin, Y., Monnet, G. ET AL. (2001a), MNRAS, 326, 23
Bacon, R., Emsellem, E., Combes, F. ET AL. (2001b), A&A, 371, 409
Bahcall, J.N., Tremaine, S. (1981), ApJ, 244, 805
Beck, T.L., Riera, A., Raga, A.C., Aspin, C. (2004), AJ, 127, 408
Binney, J., Tremaine S. (1987), Galactic dynamics, Princeton University Press
Bonnet, H., Abuter, R., Baker, A. ET AL. (2004), Msngr, 117, 17
Bower, R.G., Morris, S.L., Bacon, R. ET AL. (2004), MNRAS, 351, 63
Comte, G., Marcelin, M. (eds. 1995), ASPCS, 71
Davis, C.J., Varricatt, W.P., Todd, S.P., Ramsay Howat, S.K. (2004), A&A, 425, 981
Dougados, C., Cabrit, S., Lopez-Martin, L., Garcia, P., O’Brien, D. (2003), Ap&SS,
287, 135
Dumas, C. ET AL. in preparation
Eisenhauer, F., Schödel, R., Genzel, R. ET AL. (2003a), ApJ, 597, L121
Eisenhauer, F., Abuter, R., Bickert, K. ET AL. (2003b), Proc. SPIE, 4841, 1548
Eisenhauer, F., Genzel, R., Alexander, T. ET AL. (2005), ApJ, 628, 246
Emsellem, E., Cappellari, M., Peletier, R.F. ET AL. (2004), MNRAS, 352, 721
Förster Schreiber, N.M., Genzel, R., Lehnert, M.D. ET AL. (2006), ApJ, 645, 1062
Francis, P.J., Wills, B.J. (1999), ASPCS, 162, 363
Garcı́a, P.J.V. (2006), New Astronomy Review, 49, 590
Genzel, R., Thatte, N., Krabbe, A., Kroker, H., Tacconi-Garman, L.E. ET AL. (1996),
ApJ, 472, 153
Genzel, R. ET AL. (1997), in SINFONI a High Resolution Near-Infrared Imaging Spectrometer
for the VLT, www.mpe.mpg.de/SPIFFI/preprints/STC.210.ps.gz
Genzel. R., Pichon, C., Eckart, A., Gerhard, O.E., Ott, T. (2000), MNRAS, 317, 348
Genzel, R., Schödel, R., Ott, T. ET AL. (2003a), ApJ, 594, 812
Genzel, R., Schödel, R., Ott, T. ET AL. (2003b), Natur, 425, 934
Ghez, A.M., Duchêne, G., Matthews, K. ET AL. (2003), ApJ, 586, L127
Herbst, T.M., Beckwith, S.V.W., Glindemann, A. ET AL. (1996), AJ, 111, 2403
Kasper, M.E., Feldt, M., Herbst, T.M., Hippler, S., Ott, T., Tacconi-Garman, L.E.
(2002), ApJ, 568, 267
Kissler-Patig, M., Roth, M.M., Walsh, J.R. (2007), in Science Perspectives for 3D Spec-
troscopy, ESO Astrophysics Symposia, eds. Kissler-Patig, M., Wals, J.R., & Roth, M.M.
Springer
Krabbe, A., Genzel, R., Eckart, A., ET AL. (1995), ApJ, 447, L95
Lavalley, C., Cabrit, S., Dougados, C., Ferruit, P., Bacon, R. (1997), A&A, 327, 671
Lavalley-Fouquet, C., Cabrit, S., Dougados, C. ET AL. (2000), A&A, 356, L41
Leitherer, C., Schaerer, D., Goldader, J.D. ET AL. (1999), ApJS, 123, 3
Leonard, P.J.T, Merrit, D. (1989), ApJ, 339, 195
Levin, Y., Beloborodov, A.M. (2003), ApJ, 590, L33
Martin, P.D., Pinet, P.C., Bacon, R. ET AL. (1996), P&SS, 44, 859
McDermid, R., Bacon, R., Adam, G., Benn, C., Cappellari, M. (2004), Proc. SPIE, 5492,
822
McGregor, P.J., Dopita, M., Wood, P., Burton, M.G. (2001), PASA, 18, 41
Monreal-Ibero, A., Roth, M.M., Schönberner, D., Steffen, M., Böhm, P. (2005), ApJ,
628, L139
Müller-Sánchez, F. ET AL. (2006), A&A, accepted, astro-ph/0604317
Nagar, N.M., Falcke, H., Wilson, A.S. (2005), A&A, 435, 521
Narayan (2002), in Lighthouses of the Universe: The Most Luminous Celestial Objects and
Their Use for Cosmology, Proceedings of the MPA/ESO Conference, 405
Osterbrock, D.E. (1989), Astrophysics of gaseous nebulae and active galactic nuclei, Univer-
sity Science Books
Reid, M.J., Brunthaler, A. (2004), ApJ, 616, 872
Rutten R.G.M., Benn, C.R., Méndez, J. (eds. 2006), New Astronomy Reviews, 49, 487
Schmidt, M. (1963), Natur, 197, 1040
Schödel, R., Ott, T., Genzel, R. ET AL. (2002), Natur, 419, 694
Shu, F.H., Adams, F.C., Lizano, S. (1987), ARA&A, 25, 23
Swinbank, A.M., Smail, I., Bower, R.G. ET AL. (2005), MNRAS, 359, 401
Tecza, M., Genzel, R., Tacconi, L.J. ET AL. (2000), ApJ, 537, 178
Van Breugel, W., Bland-Hawthorn, J. (2000), ASPCS, 195
Van Zyl, L., Charles, P.A., Arribas, S., Naylor, T., Mediavilla, E., Hellier, C. (2004),
MNRAS, 350, 649
Vasconcelos, M.J., Cerqueira, A.H., Plana, H., Raga, A.C., Morisset, C. (2005), AJ,
130, 1707
Weitzel, L., Krabbe, A., Kroker, H. ET AL. (1996), A&AS, 119, 531
6. Extragalactic studies and future integral
field spectroscopy science
LUIS COLINA
In this set of lectures, I review recent observational progress on extragalactic studies using
integral field spectroscopy (IFS) techniques, highlighting the importance of IFS for the
study of the nuclear regions of nearby galaxies, of low-z active galactic nuclei (AGN)
and massive star-forming galaxies, and of high-z galaxies, including lensed quasars,
lensing galaxies and bright submillimetre galaxies. Emphasis is given to the study of
(ultra)luminous infrared galaxies as examples of low-z systems where the physical pro-
cesses relevant to the formation and evolution of galaxies can be investigated in more
detail. Research projects involving future ground-based facilities and satellites are also
briefly presented.
6.1 Introduction
The use of IFS for extragalactic studies has burgeoned over the past 10 years and is
already becoming a standard observational technique used by several groups in many
different areas. Most IFS systems (INTEGRAL, GMOS, PMAS, SAURON, SINFONI,
VIMOS, etc.) allow us to simultaneously obtain spectra covering a wide spectral range
over a wide field of view (up to 1 arcmin square for VIMOS). These instruments in their
standard configurations provide low–intermediate spectral resolution (R of 1000 to 4000)
with a relatively low angular resolution (0.5′′ to 3.0′′ ). In addition, a few IFS systems, such
as OASIS on the William Herschel Telescope and SINFONI on the Very Large Telescope
(VLT), can provide very high angular resolution (i.e. 0.1′′ ) in the optical (OASIS) and
near-infrared (SINFONI) when combined with adaptive optics (AO) systems. Finally,
multi-object integral field spectrographs (GIRAFFE on the VLT) are already available
that are leading the way to future IFS systems currently in design (see Section 6.3). A
full account of the different available IFS systems and their characteristics can be found
in M. Bershady’s contribution to this Winter School.
Studies of galaxies at low and high redshift benefit greatly from the flexibility of IFS
systems. Detailed studies on different angular scales – from the nuclear, to circumnuclear
and external regions – can only be executed efficiently with IFS. Moreover, these systems
allow us to investigate simultaneously the two-dimensional structure of many astrophys-
ically important properties such as chemical composition, stellar populations, ionization
conditions, kinematics, etc. that are directly related to the formation and evolution of
galaxies.
Among the several areas of extragalactic research covered by IFS-based studies, this
contribution will focus on the study of (ultra)luminous infrared galaxies – (U)LIRGs –
explaining in detail the impact and use of IFS in the investigation of several topics, such
as the structure of the internal dust extinction, the properties of young stellar clusters, the
presence of different ionizing sources such as hidden AGN, (extra)nuclear star-forming
regions and shocks, and the kinematic properties of the nuclear and extended ionized
gas. All these aspects are common to many other research areas, such as the study of
the nuclear regions in nearby ellipticals and spirals, of the ionization structure in low-z
AGN, of extended nebulae in low- and high-z galaxies, and of studies of high-z galaxies,
including lensed objects. Due to space limitations, these studies will not be covered in
200
Extragalactic studies and future IFS science 201
10 arcsec
[O III] Hβ
Arcsec
SAURON WFPC2HST
gas gas stars stars

Arcsec
E
Arcsec Arcsec Arcsec Arcsec
–100 +30 0 110
V [km/s] +100 0 σ [km/s] 170 –30 V [km/s] σ [km/s]
Figure 6.1. Maps of NGC 7742 taken with SAURON. The top panels show the emission line
intensity distributions of [O III] and Hβ, followed by a SAURON reconstructed image composed
of [O III], blue continuum and red continuum, and a similar image composed from HST/WFPC2
filters. The bottom row shows (from left to right) the derived gas velocity and velocity dispersion
fields, and the stellar velocity and velocity dispersion fields (figure reproduced from de Zeeuw
et al., 2002).
detail here and the reader is referred to the relevant references given in the following as
part of this introduction.
6.1.1 Stellar populations and kinematics in early-type galaxies

The formation and evolution of early-type galaxies is being investigated by studying the
fossil record (i.e. the two-dimensional kinematics and ages of the stellar population) for
a complete sample of nearby ellipticals and spirals using mostly SAURON, a dedicated
IFS system on the William Herschel Telescope (Bacon et al., 2001; de Zeeuw et al., 2002;
see Figure 6.1 for an example). On kiloparsec scales, many objects display circumnuclear
stellar kinematic twists, kinematically decoupled components, central stellar discs and
other peculiarities with almost no age differences with their host galaxy (Emsellem et al.,
2004). However, on scales of hundreds of parsecs, additional high angular resolution IFS
studies with OASIS reveal the presence of previously undetected kinematically distinct
nuclear components that also show a large fraction of young stars (McDermid et al., 2006).
On the other hand, the kinematics of the ionized gas is often decoupled from that of the
stars, is rarely consistent with coplanar circular motions, but shows large-scale coherent
motions and smooth variations (Sarzi et al., 2006). The recent extension of these studies
to late-type (Sb to Sd) spirals indicate different, more complicated features in the velocity
fields of these galaxies with respect to the early types. These features include wiggles in
the zero-velocity line, irregular distributions and ring-like structures (Ganda et al., 2006).
6.1.2 Ionization structure and kinematics in AGN

Research on low-z AGN has been well represented in IFS science over the last 10 years.
Interest in learning about the kinematics and ionization in the nuclear regions close to the
AGN engine, and in the associated starburst−AGN connection has been the main topic
202 Luis Colina
Stellar continuum Stellar velocity field
20 3.9 20 1100
10 3.6 10 1000
∆d (arcsec)
∆d (arcsec)
0 3.3 0 900
–10 3 –10 800
–20 2.7 –20 700

20 10 0 –10 –20 20 10 0 –10 –20
∆a (arcsec) ∆a (arcsec)
[O III] Emission [O III] Velocity Field
20 4.5 20 1100
10 3.8 10 1000
∆d (arcsec)
∆d (arcsec)
0 3.15 0 900
–10 2.47 –10 800
–20 1.8 –20 700

20 10 0 –10 –20 20 10 0 –10 –20
Hb Emission Hb Volocity Field
20 4000 20 1100
10 3000 10 1000
∆d (arcsec)
∆d (arcsec)
0 2000 0 900
–10 1000 –10 800
–20 0 –20 700

20 10 0 –10 –20 20 10 0 –10 –20
Figure 6.2. Intensity (arbitrary units) and velocity (in km s−1 ) maps for NGC 5033 taken with
INTEGRAL. The intensity maps correspond to the stellar continuum (4700–5840 Å wavelength
interval), and the ionized gas traced by the [O III] and Hβ emission. The star and the circle
in the velocity maps represent the location of the intensity peak in the stellar continuum and
[O III] intensity maps, respectively (figure reproduced from Mediavilla et al., 2005a).
in this area. Detailed studies of the galaxy NGC 1068, considered as the prototype of
Seyfert 2 galaxies, include early work on the biconical ionization field (Arribas et al., 1996)
and on the kinematics of the central nuclear and circumnuclear region (Garcı́a-Lorenzo
et al., 1997, 1999). Similar individual studies of other well-known nearby galaxies on scales
of hundreds of parsecs and covering a wide range of activity from LINERs to Seyfert 1
galaxies have been undertaken from the early days of IFS to the present (see Figure 6.2
for an example; Arribas and Mediavilla, 1994; Arribas et al., 1997, 1999; Mediavilla et al.,
1997, 2005a; Garcı́a-Lorenzo et al., 2001; Ferruit et al., 2004). With the advent of near-IR
AO-based IFS systems such as SINFONI, the star formation in the inner regions close to
the AGN, the properties of the molecular gas in the central 100 pc region and its relation
to the torus is starting to be investigated on spatial scales of a few to several parsecs
(Davies et al., 2006; Mueller-Sánchez et al., 2006).
Not only has work on AGN covered radio-quiet AGN, but IFS has also been used
to investigate the kinematics and ionization of the interaction of the radio jet with the
surrounding interstellar medium in the radio galaxy 3C 120 (Sánchez et al., 2004; Garcı́a-
Lorenzo et al., 2005). VIMOS on the VLT is now also being used to obtain deep IFS of
the giant nebulae with sizes in excess of 100 kpc that surround high-z radio galaxies, and
4 4
a)
2 2
∆d (arcsec)
0 0
–2 –2
–4 –4
4 2 0 –2 –4 4 2 0 –2 –4
4 4
b)
2 2
∆d (arcsec)
0 0
–2 –2
–4 –4
4 2 0 –2 –4 4 2 0 –2 –4
4 4
c)
2 2
∆d (arcsec)
0 0
–2 –2
–4 –4
4 2 0 –2 –4 4 2 0 –2 –4
Figure 6.3. Image representing six of the twenty continuum maps obtained for the lensed sys-
tem SBS 0909+532 using INTEGRAL. a) to c), left to right: Continuum central wavelengths are
8727 Å, 7233 Å, 5739 Å, 4447 Å, 4006 Å and 3487 Å. Changes in the relative brightness of the two
components as a function of wavelength are due to internal extinction within the intermediate
redshift lens galaxy (figure reproduced from Motta et al., 2002).
to establish their origin. The kinematics of the nebulae around the galaxy MRC2104-242
at a redshift of 2.49 is consistent with rotation, therefore implying a dynamical mass of
at least 3 × 1011 M⊙ in this system (Villar-Martı́n et al., 2006).
6.1.3 Interstellar medium in high-z gravitational lens systems

In the past few years IFS has been used to investigate the internal properties of both
intermediate-z lens galaxies and high-z QSOs (quasi-stellar objects) and lensed galaxies.
A precise determination of the internal extinction law in the lens galaxy of the system
SBS 0909+532, at a redshift of 0.83, has been obtained combining IFS in the optical
(Figure 6.3; Motta et al., 2002) and HST spectroscopy (Mediavilla et al., 2005b). The
IFS study of the Einstein Cross (i.e. lensed QSO 2237+0305) has allowed the detection of
the extended narrow line region (NLR) in this QSO, establishing a size of 0.7 to 0.9 kpc
(Mediavilla et al., 1998; Motta et al., 2004). Finally, the gravitational potential of the
cluster A2218 has been used as a lens magnifier to derive the resolved kinematics of
the ionized gas in a lensed spiral galaxy at a redshift of 1.034, demonstrating that this
galaxy lies close to the mean Tully–Fisher relation of present-day spirals (Swinbank et al.,
2003). Further analysis to investigate the emission line properties of several galaxies
at intermediate redshifts (z ∼ 1) have recently been reported (Swinbank et al., 2006a)
applying the same technique, and exploiting the gravitational lens effect due to low-z
massive clusters.
204 Luis Colina
6.1.4 Formation and evolution of high-z galaxies
Over the last decade deep multiwavelength surveys have identified a large diversity of
galaxy populations in the high-z Universe. The morphological evolution of galaxies and
the contribution of these different populations to the star formation history of the uni-
verse is being investigated mostly through deep high-angular resolution imaging and
stellar population modelling. However, our understanding of how galaxies of different
mass are assembled over cosmic time is comparatively very poor. In order to progress
in this area a better knowledge of different physical processes – such as gravitational
interactions due to mergers, feedback from star formation and AGN, exchange and
loss of angular momentum, and metal enrichment of the interstellar and intergalactic
medium – is needed. Due to the complex structure in a large fraction of high-z galaxies,
IFS has a clear advantage over pure imaging or long-slit spectroscopy to investigate the
formation and evolution of galaxies (Colina et al., 2005; Förster Schreiber et al., 2006;
Swinbank et al., 2006b).
The Tully–Fisher relation (Tully and Fisher, 1977) has been studied in a sample of
intermediate redshift (0.4 < z < 0.75) galaxies using the first operational multi-object IFS
system, FLAMES/GIRAFFE on the VLT (Flores et al., 2006). Of the galaxies surveyed
only 35% are rotating discs showing the same slope, zero point and scatter as the low-z
TF relation. The rest of the galaxies do not follow the TF relation, indicating a strong
dynamical evolution (minor or major mergers, merger remnant, or radial gas motions)
traced by the observed two-dimensional complex kinematics (Flores et al., 2006).
The new optical and near-IR IFS systems in 10 m-class telescopes are also being used to
obtain the two-dimensional ionization and kinematic structure of galaxies at redshifts of 2
and above (Förster Schreiber et al., 2006; Swinbank et al., 2006b). Massive submillimetre
galaxies, such as SMM J14011+0252 at a redshift of 2.565 (Tezca et al., 2004) and SMM
J16365+5047 (N2 850.4) at a redshift of 2.385 (Figure 6.4; Swinbank et al., 2005), are
already been studied. The presence of large peak-to-peak velocity differences, decoupled
stellar and gas structures and evidence for AGN have already been detected (Tezca et al.,
2004; Swinbank et al., 2005). Many of these systems comprise two or more dynamical
subcomponents with velocity offsets of 180 km s−1 across scales of several kpc (Swinbank
et al., 2006b), suggesting that these systems are analogous to low-z Ultra Luminous
Infrared Galaxies (ULIRGs; Colina et al., 2005). A more extensive investigation of the
Hα rest-frame emission in a sample of UV-selected galaxies at an average redshift of
2.2 has already been done using SINFONI on the VLT (Förster Schreiber et al., 2006).
Spatially resolved velocity fields with peak-to-peak velocities of 40 to 410 km s−1 over
scales of 10 kpc have been identified in 60% of the galaxies. These velocity fields are
generally consistent with orbital motions, although in only four galaxies (i.e. 35% of the
sample) are they well described by rotating disks (Förster Schreiber et al., 2006).
6.2 IFS studies of low-z ULIRGS

6.2.1 Astrophysical relevance of (Ultra) Luminous Infrared Galaxies
The importance of luminous and ultraluminous (LIRGs, LIR = L[8–1000 µm] = 1011 –
1012 L⊙ , and ULIRGs LIR ≥ 1012 L⊙ , respectively) has been recognized since their dis-
covery (Rieke and Low, 1972) and the detection of large numbers by the IRAS satellite
(Soifer et al., 1987). The process of interaction/merging (Farrah et al., 2001; Bushouse
et al., 2002, and references therein) with accompanying starbursts (Lutz et al., 1998) that
produces (U)LIRGs, appears to be an important stage in the evolution of galaxies. The
transformation of gas-rich (Solomon et al., 1997) moderate-mass (< m∗ ) disk galaxies
HST ACS/NICMOS I814H160 colour image a) IRTF Ho + NICMOS H160 b)
1 1
Arc seconds
Arc seconds
0 0
E E
–1 –1
N N
–1 0 1 –1 0 1
HST NICMOS H160 + Ho velocity field c) D d)
1 A 1
C
Arc seconds
Arc seconds
0 0
B
E E
–1 –1
0 kms–1 +270 N N
HST ACS I814 + GMOS IFU Lya
–1 0 1 –1 0 1
Arc seconds Arc seconds
Figure 6.4. a) Image of N2 850.4 combining HST ACS and NICMOS imaging. The image shows
a complex morphology, with at least three distinct components separated by 1 arcsec (8 kpc)
in projection. b) Infra-Red Telescope Facility (IRTF) Hα narrow-band image with the contours
from the NICMOS H-band image overlaid. This Hα narrow-band image shows a diffuse halo of
material distributed asymmetrically around the galaxy. c) The velocity field derived from UIST
IFS observations of the Hα emission line overlaid on the NICMOS H-band image. Components
A and B are separated by 50 ± 50 km s−1 whilst there is a velocity difference of 270 ± 50 km s−1
between components A and C. d) HST ACS I-band image with the Lyα intensity from the
GMOS IFS overlaid as contours (figures reproduced from Swinbank et al., 2005).
(Colina et al., 2001, 2005) into moderate-mass ellipticals through the process of merging
(Genzel et al., 2001; Tacconi et al., 2002; Dasyra et al., 2006) suggests a physical process
for forming early-type galaxies in the early Universe, such as those recently detected at
redshifts of 2–3 (Franx et al., 2003; Cimatti et al., 2004; Förster Schreiber et al., 2004;
van Dokkum et al., 2004).
Deep submillimetre and radio surveys (Chapman et al., 2003a, 2003b, and references
therein) are detecting a population of objects, the so-called sub-millimetre galaxies
(SMGs), that appear to be forming stars at rates of up to 1000 M⊙ yr−1 (i.e. up to
10 times more than ULIRGs), and that, according to their morphology (Chapman et al.,
2003b), luminosity and spectral energy distribution (Egami et al., 2004; Frayer et al.
2003), seem to be the high-z counterparts of the low-z luminous (LIRG) and ultralumi-
nous (ULIRG) infrared galaxies (for a review, see Blain et al., 2002). Mid-IR satellites
(Spitzer Space Telescope as well as millimetre interferometers, IRAM, ALMA) continue
to detect high-z LIRGs and ULIRGs in large quantities, as already shown by the data
taken with Spitzer (Charmandaris et al., 2004; Egami et al., 2004; Le Floc’h et al., 2004).
The majority of these galaxies at z < 1 are in the LIRG class, and they make a significant
contribution to the star formation density at 0.5 < z < 2 (Pérez-González et al., 2005).
Studies of the evolution of the morphological types of galaxies as a function of redshift
indicate that, as in low-z ULIRGs, mergers have played a major role in the shaping
of present-day galaxies. These studies (Conselice et al., 2003, 2005) conclude that the
fraction of galaxies consistent with undergoing a major merger increases with redshift
206 Luis Colina
HST/WFPC2 F814W
5 C'
C1
B21
0
D3
–5 B2
B1
B16
–5
Ha Hb
–5
5 0 –5 5 0 –5 5 0 –5
Figure 6.5. Continuum and emission line images for NGC 3690 obtained with INTEGRAL.
The contours represent the red stellar continuum at 6460 Å. The HST/WFPC2 I-band (F814W)
image is shown for comparison, with several regions of interest marked on it. Component B2
dominates the stellar continuum emission while components C1 and B1 are the strongest emis-
sion line sources (figure reproduced from Garcı́a-Marı́n et al., 2006).
for all galaxies, and that at redshifts above 2, more than 80% of the stellar mass is in
objects with peculiar morphology. There is also a predominance of irregular and complex
morphologies in high-z luminous SMGs, suggesting that major mergers are common in
these dusty star-forming galaxies (Chapman et al., 2003b). In addition, galaxies less
affected by dust – such as the Lyman break galaxies (LBGs) and UV-selected, high-z
star-forming galaxies – do often show irregular and distorted morphologies (Erb et al.,
2003, 2004; Giavalisco et al., 1996), indicating that interaction/merging is also involved
in these galaxies.
6.2.2 Structure of the stellar and ionized gas distribution

The stellar and ionized gas distribution is traced by line-free stellar continuum images
and by the light of several optical emission lines, respectively (see Figures 6.5 and 6.6 for
WFPC2 F814W Continuum Ha Continuum Hb
4 4 4
R2
Nn
2 2 2
0 0 0
–2 –2 –2
R1
Ns
–4 –4 –4
4 2 0 –2 –4 4 2 0 –2 –4 4 2 0 –2 –4
[OIII] Ha Hb
4 4 4
2 2 2
0 0 0
–2 –2 –2
–4 –4 –4
4 2 0 –2 –4 4 2 0 –2 –4 4 2 0 –2 –4
Figure 6.6. INTEGRAL images of the ionized gas and stellar light distribution in the central
regions of IRAS 12112+0305 as traced by different emission lines and by continuum windows
close to Hβ and Hα. The observed [O III] and Hβ emission is dominated by an extinction-free
extranuclear star-forming region. The HST I-band (F814W) image is shown for comparison
(figure reproduced from Colina et al., 2000).
examples). The wavelength dependence of the optical and near-IR stellar light distribu-
tion is mostly produced by the absorption due to the large amount of dust towards the
nucleus in these galaxies. A clear example of high nuclear extinction is shown in IRAS
12112+0305 where the red bright, compact southern nucleus (Ns ) appears as a weak
source in the Hβ-continuum light (Figure 6.6) seen through an extinction of about eight
magnitudes in the visual (Colina et al., 2000). In other galaxies, such as Arp 220 and
IRAS 17208−0014, the presence of a strong central dust lane obscures the true nucleus in
the optical. As a result, the optical nucleus, identified as the high surface brightness peak,
does not spatially coincide with the near-IR nucleus but is about 1 kpc apart in these
galaxies (Arribas et al., 2001; Arribas and Colina, 2003). In several ULIRGs investigated,
the internal extinction produces a wavelength dependence of the relative brightness of the
heavily obscured nuclear regions with respect to the less obscured external van Dokkum
regions of the host galaxy while keeping the overall stellar light distribution, even in
interacting pairs.
The observed ionized gas and stellar distributions show decoupled structures with
different morphologies in many (U)LIRGs. These differences are due to the presence of
extranuclear gas ionized by a central active galactic nucleus (Mrk 273; Colina et al., 1999;
Arp 299; Garcı́a-Marı́n et al., 2006), the presence of bright, tidally induced extranuclear
star-forming regions along the tidal tails (IRAS 14348−1447; Monreal-Ibero, 2004) or
far away (9 kpc) from the nucleus of the galaxy (IRAS 12112+0305; Colina et al., 2000)
and the presence of circumnuclear gas ionized by a nuclear, dust-enshrouded starburst
(Arp 220; Arribas et al., 2001). In addition, the complex, non-axisymmetric structure
of the internal extinction in these galaxies produces additional wavelength-dependent
effects such as changes in the relative brightness of the nuclear regions with respect
208 Luis Colina
to less absorbed star-forming regions (IRAS 08572+3915; Arribas et al., 2000), or to
extranuclear regions located several kpc away from the nucleus (IRAS 12112+0305).
The ionized gas distribution traced by the emission lines shows a different structure
in some of the galaxies. Since these emission lines trace different ionization mechanisms
(shocks, young stars, AGN), their two-dimensional light structure represents the spatial
distribution of these different ionizing sources within a given galaxy. This is clearly seen
in the two nearest galaxies studied so far, Arp 299 and Arp 220. The peaks of the [O I]
λ6300 Å, [O III] λ5007 Å, Hα and [S II] λλ6717,6731 Å emitting regions in Arp 299 are not
spatially coincident and present substructure even on scales of 200 pc or less (Figure 6.5
and Garcı́a-Marı́n et al., 2006). The high surface brightness plumes in Arp 220, detected
up to 4 kpc from the nucleus, present a complex ionization structure as identified by the
changes in the [N II]/Hα emission line ratio measured along the plumes (Colina et al.,
2004).
6.2.3 Star formation and dust extinction

The distribution of dust, and consequently its absorption effects, presents a complex
morphology and a wide range of extinction values measured using optical and/or near-
IR hydrogen recombination lines. Within the same galaxy, star-forming regions almost
free of dust (i.e. ≤ 1 V magnitude) coexist with heavily obscured nuclear starbursts
(≥ 3 magnitudes in V) on scales of hundreds to thousand of parsecs (Figure 6.7; Garcı́a-
Marı́n et al., 2006). If corrected by the two-dimensional extinction map, the Hα light
distribution agrees with the less-absorbed ionized gas distribution traced by the near-IR
Paα emission line (Figure 6.7).
Dust absorption effects and their impact in the observed extranuclear versus nuclear
star formation are clearly seen in many ULIRGs, such as IRAS 12112+0305 (Figure 6.6;
Colina et al., 2000). The stellar main body of this galaxy is concentrated in three regions:
the two nuclei (Nn and Ns ) and a third region (R2 hereafter) about 3′′ north of Ns .
However, the observed dominant line-emitting region (R1) does not coincide with any
of the nuclei identified above but is located 5′′ (i.e. 7.5 kpc) east of Ns . Moreover, the
optically dominant nucleus of the galaxy (Ns ) appears as the faintest Hα emission source.
These effects are due to the relative extinction of the different regions. While region R1
has an extinction of only 0.9 magnitudes in the visual, in the southern nucleus (Ns ) a
visual extinction of 7.8 magnitudes is measured from the Hα/Hβ line ratio (Colina et al.,
2000). Thus, although Ns is a minor contributor to the observed Hα emission, the
reddening-corrected flux emanating from this nucleus dominates the overall Hα lumi-
nosity. The corresponding star formation rate would amount to about 74 M⊙ yr−1 for a
Salpeter initial mass function (IMF), with mass limits of 0.1 and 100 M⊙ (Colina et al.,
2000). In many of the (U)LIRGs the extinction is so high (> 10 visual mag) in the
nuclear (IRAS 17208−0014; Arribas and Colina, 2003) or circumnuclear regions (NGC
7469, Dı́az-Santos et al., 2007) that an important fraction of the star formation is
not detected in the optical, hence the importance of future near-IR and mid-IR IFS
studies.
6.2.4 Detection of AGN

Traditionally, the level of activity in galaxies has been obtained through diagnostic
diagrams (BPT diagrams, after Baldwin, Phillips and Terlevich, 1981) based on the
ratios of different optical emission lines. These studies have been restricted to the nuclei
of ULIRGs or to extended regions along a given direction (Veilleux et al., 1992). Some
ULIRGs, such as Mrk 231, I Zw 1 or Mrk 273, are known to host a powerful AGN. These
AGN already show the spectral characteristics of Seyfert 1 (broad emission lines) or
a) b)
5 C'
C1
B21
0
D3
–5 B2
B1
B16 HST F190N HST F160W
1.50
c) d)
5 1.25
1.01
E(B–V)
0 0.77
0.53
–5 0.29
0.05
5 0 –5 5 0 –5
Figure 6.7. a) Panel showing the galaxy NGC 3690 the HST/NICMOS continuum-subtracted
Paα emission; b) the HST/NICMOS H-band continuum; c) the INTEGRAL Hα extinction
corrected emission and d) the extinction spatial distribution obtained with the Hα/Hβ line
ratio. The contours represent the red stellar continuum at 6460 Å as obtained with INTEGRAL
(figure taken from Garcı́a-Marı́n et al., 2006).
Seyfert 2 (strong high excitation [O III] λ5007 Å emission line) nuclei. In most ULIRGs,
however, the presence of a dust-enshrouded AGN is not evident in the optical, based
only on the nuclear spectrum. Integral field spectroscopy allows us to investigate the
two-dimensional ionization structure of the nuclear and circumnuclear gas searching for
traces of optically dust-enshrouded AGN. These types of AGN can be traced by the pres-
ence of extended extranuclear high-excitation ionized regions similar in size and structure
to those detected in nearby AGN. Some ULIRGs, such as Mrk 273 (Colina et al., 1999)
and IRAS 11087+5351 (Garcı́a-Marı́n et al., in preparation), show the presence of highly
ionized [O III] emitting regions at distances of several kiloparsecs from the nucleus of the
galaxies. Some other galaxies show a well-defined ionizing cone morphology, as expected
from the standard AGN model (Urry and Padovani, 2000 and references therein). One
of these cones, detected in the galaxy NGC 3690, is in the process of merging with IC
694 and forms the luminous infrared system Arp 299 (Figure 6.8; Garcı́a-Marı́n et al.,
2006). According to the standard diagnostic diagrams, three different types of ionization,
HII-LINER and AGN-like, are identified. The line ratio maps clearly show the presence of
an extranuclear Seyfert-like nebula with a well-defined conical morphology, with its apex
in region B1, and located at a projected distance of 1.5 kpc from its apex. This highly
ionized conical structure in NGC 3690 is also detected at larger distances from B1 in the
210 Luis Colina
Figure 6.8. Diagnostic diagram and line ratio maps derived for the LIRG Arp 299 (IC 694
+ NGC 3690) using INTEGRAL with the SB3 bundle. a) [O III] λ5007 Å/Hβ versus [O I]
λ6300 Å/Hα diagnostic diagram with all the data displayed. The symbol code is as follows:
triangles represent values within the ionization cone, and at distances of less than 2 kpc from
region B1; stars represent the region within the cone at distances of 2–4 kpc from B1; crosses
mark the interface region between the two galaxies that form the Arp 299 system (IC 694 and
NGC 3690); and filled circles represent the regions in the two galaxies (defined as the regions
within the isocontours). b)–c) [OIII] λ5007 Å/Hβ and [O I] λ6300 Å/Hα maps, respectively. The
nuclei of galaxies IC 694 and NGC 3690 are identified with regions A and B1, respectively. The
white line delineates the ionization cone with its apex in B1 (figure reproduced from Garcı́a-
Marı́n et al., 2006).
[O I] λ6300 Å/Hα map (Figure 6.8). However, the ionization associated with the outer
regions of the cone is LINER-like rather than Seyfert, indicating that within the cone
there are two ionization regimes well separated spatially. Therefore, the two-dimensional
ionization maps strongly support photoionization by a radiation cone escaping from a
central dust-enshrouded AGN source located in B1. This AGN has also been recently
identified at other wavelengths (Della Ceca et al., 2002; Ballo et al., 2004; Gallais et al.,
2004).
Figure 6.9. Spatial distribution of the spectra (centred on the Hα-[N II]) in the circumnuclear
region of Arp 220 obtained using the bundle SB2 of INTEGRAL (figure reproduced from Arribas
et al., 2001).
6.2.5 Arp 220: the nearest ULIRG

The kinematic structure of the ionized gas in the central one to two kiloparsec region, as
well as in the extended, several tens of kiloparsec regions, can only be investigated effi-
ciently using IFS with different angular resolutions from sub-arcsec to several arcsec. IFS
systems like INTEGRAL and PMAS offering bundles of fibres with different diameters
from 0.5′′ to 3.0′′ are ideal for these types of studies. So far, because of the low surface
brightness of the regions, a detailed study of only the nearest ULIRG, Arp 220, covering
a field of view of 75.0′′ by 40.0′′ (i.e. 25 by 18 kpc) has been performed (Arribas et al.,
2001; Colina et al., 2004).
(Circum)nuclear regions: rotation and starburst-driven outflows

The complex gas kinematics in the (circum)nuclear one to two kiloparsec region of Arp
220 is already suggested by the dramatic changes detected in the profiles of the Hα-[N II]
lines on scales of less than a few hundred parsec using IFS (Figure 6.9; Arribas et al.,
2001). To disentangle the different components of the velocity field, a two-dimensional,
multi-Gaussian line decomposition analysis has to be performed in such complex emission
line profiles (for representative examples, see Figure 6.10; for details, see Arribas et al.,
2001). Three kinematically distinct and spatially extended velocity components have
been identified in the ionized gas. One narrow (270 km s−1 FWHM; full-width-at-half-
maximum) component (A in Figure 6.11) is detected nearly everywhere over the central
14 × 12 arcsec2 region (i.e. 5.2 kpc × 4.5 kpc). The velocity field of this component with
a peak-to-peak velocity of 450 km s−1 is interpreted as gas located in a rotating disc
inclined by about 45 degrees with respect to the line of sight and with the spin axis
along the SE–NW direction (position angle 135◦ ). The inclination-corrected rotational
212 Luis Colina
Figure 6.10. Some examples showing the complex spectra in the nuclear regions of Arp 220
and the fitted components indicating the presence of spatially separated velocity components.
Numbers represent the fibre as displayed in the previous figure (reproduced from Arribas et al.,
2001).
Figure 6.11. Velocity field and spatial distribution of the three kinematically distinct gaseous
components in the circumnuclear regions of Arp 220. These velocity components are associated
with rotational motions (A) and radial flows (figure reproduced from Arribas et al., 2001).
velocities imply a dynamical mass (Mdyn ) of 3.5 × 1010 M⊙ within a radius of 1.5 kpc.
This relatively high value indicates a large mass concentration in the nuclear region of
Arp 220, as already inferred by the presence of 5 × 109 M⊙ of molecular gas in a nuclear,
0.5 kpc size disc (Scoville et al., 1997).
A broad velocity component (B in Figure 6.11) with an average width (FWHM) of
815 km s−1 is identified over an extended region of 1 kiloparsec in diameter, at about
600 pc northwest of the dust-enshrouded nucleus, and blueshifted by 300 km s−1 with
respect to the system velocity. A third, high-velocity component (C in Figure 6.11) with
peak-to-peak velocities of 1000 km s−1 is detected mostly southeast of the nucleus and
orientated along the kinematic minor axis of the rotating component (for details, see
Arribas et al., 2001). The two-dimensional distribution and kinematics of components
B and C are consistent with a bipolar gas outflow characterized by an opening angle
of about 90 degrees, and orientated perpendicularly to the nuclear rotating disc of gas
traced by component A. These overall gas kinematics agree with the basic predictions
of the starburst-driven galactic wind scenario proposed by Heckman et al. (1990) for
galaxies such as (U)LIRGs with powerful nuclear starbursts.
Extended nebulae: tidally induced motions

The extended nebula around Arp 220 covering a region of 28 kpc × 15 kpc has been
mapped entirely using a low angular (2.7′′ arcsec diameter fibres) resolution mode, and
integrating for a total of close to 11 hours in three different pointings (for the final
Hα+[N II] spectra, see Figure 6.12a). These data have produced the first two-dimensional
velocity field of the entire ionized nebula in Arp 220 as well as an ionization map of the
nebulae and of the stellar light distribution on the same linear scales (see Figures 6.12b
and 6.13; Colina et al., 2004).
The two-dimensional velocity field shows its largest gradients (50 km s−1 kpc−1 ) and
deviations (+280 km s−1 to −320 km s−1 ) from the systemic velocity in the low surface
brightness ionized gas associated with the dust lane, the south-eastern stellar envelope
and the north-western broad stellar tidal tail at distances of up to 7.5 kpc from the nucleus
(Figure 6.13). Therefore, the velocity field of the extended nebula at distances of a few
to several kpc from the nucleus and with projected peak-to-peak velocity deviations
of 600 km s−1 is associated with well-defined stellar structures induced by the merger,
i.e. it is dominated by tidally induced flows and has no clear connection with a cen-
tral starburst-driven galactic wind. Moreover, the kinematics of the nebula, with average
velocities of +8 ± 79 km s−1 and −79 ± 58 km s−1 for the E and W lobe, respectively, are
to first order in agreement with the predictions of the tidally induced scenario (McDowell
et al. 2003), according to which the gas condensations in the lobes could be the products
of the merger that are animated by a combination of continuing expansion and residual
rotation from the original rotation of the two discs involved in the merger. This scenario
has to be explored further by measuring the two-dimensional velocity structure of the gas
along the rims of the lobes with better signal-to-noise spectra. Low angular resolution
integral field spectrographs mounted on 10 m-class telescopes would be ideal for estab-
lishing the role of starburst-driven winds and of merger-induced motions on different
linear scales and at different phases in the evolution of the interaction/merger process of
galaxies.
6.2.6 Merger evolution and ionized gas kinematics

The overall kinematic characteristics of the extended nebulae have been investigated with
IFS on a sample of eight ULIRGs (Colina et al., 2005). The sample, although small in
size, is representative of ULIRGs, and covers a wide range of morphologies representing
different phases of the merging process from well-separated interacting pairs (i.e. IRAS
08572+3915, IRAS 12112+0305), to double-nucleus galaxies (angular separation less than
one arcsec, i.e. Arp 220, Mrk 273) and single nucleus (i.e. IRAS 17208−0014). Also, the
full range of nuclear activity from H II and LINERs to Seyfert 1 and 2 galaxies is
covered. Finally, as defined by the ratio of the 25 µm and 60 µm IRAS bands, both warm
(f25 /f60 ≥ 0.2) and cool (f25 /f60 < 0.2) ULIRGs are included in the sample.
Velocity fields
The velocity fields of the ionized gas in the ULIRGs investigated with IFS show a complex
structure (see Figure 6.14), inconsistent in general with large, kiloparsec-scale, ordered
motions indicating that rotational motions do not dominate the velocity field. Only one
galaxy (IRAS 17208−0014) shows a velocity field consistent with that of a rotating
214 Luis Colina
Figure 6.12. a) Spatial distribution of the spectra around the redshifted Hα–[N II] spectral
region over the entire 75′′ × 40′′ extended nebulae in Arp 220. Spectra taken with INTEGRAL
using the low angular resolution bundle (SB3 with 2.′′ 7 diameter fibres) and three different point-
ings. b) Image of the stellar light distribution in the central regions and faint extended envelope
of Arp 220 as traced by a line-free narrow-band red continuum. c) Image of the ionized gas
distribution in Arp 220 as traced by the [N II] λ6484 Å emission line. The bright circumnu-
clear region, the elongated plumes and the extended diffuse lobes are easily identified. Note the
clear differences in the ionized gas distribution with respect to the stellar light presented above
(figures reproduced from Colina et al., 2004).
Integral
5800
10 (a) Arp 220
Velocity field
5600
0
∆ d (arcsec)
kms–1
–10 5400
–20 5200
–30 5000
40 20 0 –20
∆ a (arcsec)
Integral
5800
10 (b) Arp 220
Velocity field
5600
0
∆ d (arcsec)
kms–1
–10 5400
–20 5200
–30 5000
40 20 0 –20
∆ a (arcsec)
Figure 6.13. Velocity field of the extended ionized gas in Arp 220 obtained from the [N II]
λ6584 Åemission line. The contours superimposed represent the [N II] emission (a), and the
stellar light distribution, respectively (b). Orientation as in Figure 6.12 (figure reproduced from
Colina et al., 2004).
disc of about 3 kpc in size (Arribas and Colina, 2003), while one of the three velocity
components identified in the central regions of Arp 220 (see Section 6.2.5, Circum-nuclear
regions and Arribas et al., 2001) could also trace a nuclear rotating disc, already detected
in molecular gas (Downes and Solomon, 1998). In addition to IRAS 17208−0014 and
Arp 220, there are other galaxies (e.g. IRAS 08572+3915) in which the major photometric
and kinematic axes lie close in projection and could be considered as candidates for
rotating systems (Arribas et al., 2000). For the rest of the sample (seven galaxies in
six ULIRGs), the largest velocity variations, characterized by peak-to-peak differences
of 200–400 km s−1 , do not show a convincing case for rotation on scales of a few to
several kpc. Higher angular resolution AO-based IFS is needed to elucidate the presence
of rotation in the inner regions (≤ 1–2 kpc) of ULIRGs in general. Evidence for inward
flows (IRAS 15206+3342; Arribas and Colina, 2002), outflows associated with nuclear
starbursts (Arp 220; Arribas et al., 2001), flat velocity fields (Mrk 273, IRAS 15250+3609)
and motions probably associated with tidal tails (IRAS 12112+0305, IRAS 14348−1447;
Monreal-Ibero, 2004) and with the merger process (Arp 220; see Section 6.2.5.2) are also
present in these systems. The sample is still small but the results indicate that tidally
induced motions dominate the velocity field of the extended ionized nebula on kiloparsec
scales. However, there is so far no clear trend between the observed velocity field and the
phase of the interaction/merging process traced by the stellar structure, i.e. intermediate
phases identified with pairs (i.e. IRAS 12112+0305, IRAS 14348−1447), more advanced
216 Luis Colina
MRK 273
5 500 450
375 225
kms–1
0 250 0
125 –225
–5
0 –450
IRAS 05572+3915
5 150 200
113 100
kms–1
0 75 0
38 –100
–5 0 –200
Arp 220
5 400 300
300 150
kms–1
0 200 0
100 –150
–5
0 –300
IRAS 15208+4318
5 150 200
113 100
kms–1
0 75 0
38 –100
–5
0 –200
IRAS 17208–0144
5 200 250
150 125
kms–1
0 100 0
50 –125
–5 0 –250
IRAS 15250+3808
5 300 200
225 100
kms–1
0 150 0
75 –100
–5 0 –200
IRAS 14348–1447
5 300 300
225 150
kms–1
0 150 0
75 –150
–5 –300
0
IRAS 12112+0306
5 300 300
225 150
kms–1
0 150 0
75 –150
–5
0 –300
5 0 –5 5 0 –5 5 0 –5 5 0 –5 5 0 –5
Figure 6.14. Panel presenting the IFS-based kinematics of a sample of low-z ULIRGs in dif-
ferent evolutionary states. The stellar and ionized gas distribution are traced by the optical
continuum (second column from the left), and Hα line (centre column). The kinematic proper-
ties of the ionized gas are represented by the velocity dispersion map (second column from the
right) and the velocity field (right column). The Hα contours are superposed on the velocity
dispersion and velocity field maps. The near-IR (NICMOS F160W filter) HST imaging is also
shown on the left side for comparison (figure reproduced from Colina et al., 2005).
phases represented by double nuclei (i.e. Arp 220, Mrk 273) and evolved single-nucleus
galaxies (i.e. IRAS 17208−0014, IRAS 15250+3609).
Analysis of a larger sample of ULIRGs, also including the lower luminosity LIRGs, is
needed in order to characterize the gas kinematics and to establish the role and impor-
tance of starburst-driven winds, tidal-induced flows and ordered rotational motions as
the interaction/merger process evolves.
Velocity dispersion maps

The two-dimensional velocity dispersion of the ionized gas is in general obtained as 0.425
times the full width at half maximum (FWHM) of the Gaussian profile that best fits
the Hα emission line, after subtracting in quadrature the instrumental resolution and
correcting for redshift (see Figure 6.14 for final results).
There are several important results worth mentioning. In the three ULIRGs iden-
tified as pairs of interacting galaxies (i.e. IRAS 08572+3915, IRAS 12112+0305 and
IRAS 14348−1447) the velocity dispersion peak coincides with the dominant near-IR
nucleus (IRAS 08572+3915 NW, IRAS 12112+0305 SW and IRAS 14348−1447 SW).
However, the velocity dispersion of the gas associated with the secondary nuclei is either
indistinguishable (IRAS 08572+3915 SE), or even lower (IRAS 12112+0305 NE and
IRAS 14348−1447 NE), than that of the surrounding diffuse extranuclear gas. Advanced
interacting/merging systems, identified as double-nucleus within 1.0′′ and single-nucleus
galaxies, do show different behaviours. There are ULIRGs such as IRAS 17208−0014 in
which the peak of the velocity dispersion is offset with respect to the optical nucleus but
coincides with the near-IR nucleus. However, other advanced mergers (Arp 220, IRAS
15206+3342, IRAS 15250+3609) do have the peak of the velocity dispersion offset from
the near-IR nucleus by up to 2.0′′ to 3.0′′ (i.e. up to 3–5 kpc). Finally, there is one ULIRG
(Mrk 273) in which the velocity dispersion in the nuclear region is indistinguishable from
that of the gas ionized by the central AGN and located at about 3 kpc (i.e. 4.0′′ ) from
the nucleus.
In summary, the nucleus of the brightest galaxy in binary systems usually has the
highest velocity dispersion and therefore seems to trace the dynamical mass of the galaxy.
However, the high-velocity dispersions of the extranuclear gas, typically between 70 and
200 km s−1 , seem to be associated with tidally induced flows and do not in general trace
mass concentrations in these systems but extend shocks, as will be discussed in the next
section.
6.2.7 Large-scale shocks as ionization sources in extended nebulae

The ionized gas distribution in ULIRGs is characterized by the presence of compact,
high-surface-brightness regions and of well-resolved, extended nebulae with typical sizes
of between 5 to 15 kpc. In general, the ionization conditions in these extended nebulae
are characteristic of LINERs, as obtained from the diagnostic diagrams involving some of
the strongest optical emission lines (Figure 6.15). The most likely mechanism to explain
the range of observed line ratios is the presence of large-scale, high-velocity shocks with
velocities of 150–500 km s−1 . Ionization by an AGN or nuclear starburst is in general less
likely (see Figure 6.15; Monreal-Ibero et al., 2006). High-speed flows with typical peak-
to-peak velocities of 200–400 km s−1 have been detected in the tidal tails and extranuclear
regions of ULIRGs on scales of few to several kpc away from the nucleus and almost inde-
pendent of the dynamical phase of the merger (Section 6.2.6, Velocity fields and Colina
et al., 2005). Moreover, the presence of highly turbulent gas, as identified by large-velocity
dispersions of 70–200 km s−1 in extranuclear regions (see Section 6.2.6, Velocity disper-
sion maps) also supports the scenario of fast shocks as the main ionization mechanism
218 Luis Colina
Figure 6.15. [N II] λ6584 Å/Hα versus [S II] λλ6717,6731Å/Hα ratios. This figure can be
divided into several regions. In the left bottom corner is located the region occupied by typical
H II regions while the top right corner of the plot is the locus for a typical LINER-like spectrum.
Different symbols represent values for different galaxies. Values for the fibres associated with the
circumnuclear region are indicated with filled symbols, while those for the other fibres appear
with open symbols. Models of Dopita and Sutherland (1995) for shocks have been superposed.
Those without precursor are indicated with solid lines, while those with precursor are plotted
using dashed lines. At the beginning of each line the magnetic parameter Bn−1 / 2 (µG cm3 / 2 )
is shown. Shocks velocity range from 150 to 500 km s−1 for models without precursor and from
200 to 500 for models with precursor. The long-dashed line indicates the [N II] λ6584 Å/Hα
versus [S II] λλ6717,6731 Å/Hα values predicted by photoionization by a power-law model for
a dusty cloud at ne = 103 cm−3 and Z = Z⊙ (Groves et al., 2004). The dotted line indicates the
locus for an instantaneous burst model of 4 Myr, Z = Z⊙ , IMF power-law slope of −2.35 and
Mu p = 100 M⊙ (Barth and Shields, 2000); dust effects have been included and ne = 103 cm−3
(figure reproduced from Monreal-Ibero et al., 2006).
in these regions. Further evidence for the dominant role of shocks in the ionization of these
regions comes from the positive correlation measured between the velocity dispersion and
ionization status of the ionized gas in the extranuclear regions for several galaxies. Figure
6.16 represents such a correlation using the [O I] λ6300 Å/Hα line ratio, i.e. the most
reliable diagnostic ratio to detect ionization by shocks (Dopita and Sutherland, 1995),
and where each data point represents the value for a specific spectrum (fibre), i.e. differ-
ent position in the extranuclear nebula but excluding the circumnuclear region, obtained
with IFS. The fact that the extranuclear gas in a sample of galaxies covering different
dynamical phases of the merging process follows a well-defined relation between the line
ratio, and the velocity dispersion reinforces the scenario that tidal-induced shocks are the
dominant ionization source on large scales (> 2–3 kpc). However, the fact that the ioniza-
tion and velocity dispersion of the ionized gas varies from galaxy to galaxy indicates that
Figure 6.16. Relation between the shock-tracer [O I] λ6300 Å/Hα line ratio and the velocity
dispersion for the extended, extranuclear gas regions in ULIRGs. Different symbols represent
different galaxies. The dashed horizontal line marks the frontiers between the H II region and
LINER-type ionization (figure reproduced from Monreal-Ibero et al., 2006).
other mechanisms – such as ionization by a central AGN or by a distributed star forma-

tion, and turbulence due to AGN or starburst-related flows – could also be relevant in
some ULIRGs. Therefore, the universality of the observed relation is not yet established
and needs further investigation with a larger sample of (U)LIRGs.
6.2.8 (U)LIRGs and high-z star-forming galaxies

As for the low-z sample of ULIRGs investigated so far, a large fraction of high-z (> 2) star-
forming galaxies present the morphological features characteristic of interacting/merging
systems in different phases of their evolution. There is a predominance of irregular and
complex morphologies in luminous SMGs, suggesting that major mergers are common for
these galaxies (Chapman et al., 2003b). On the other hand, LBGs and UV-selected, high-
z star-forming galaxies do often show irregular and distorted morphologies (Giavalisco
et al., 1996; Erb et al., 2003, 2004), suggesting that interaction/merging is also involved
in these galaxies. A recent quantitative analysis of the morphology of galaxies detected
in the WFPC2 and NICMOS Hubble Deep Field North concludes that the fraction of
galaxies consistent with undergoing a major merger increases with redshift for all galaxies,
with the highest fraction (40–50%) of objects identified as LBGs (Conselice et al., 2003).
Moreover, at redshifts above 2 more than 80% of the stellar mass is in objects with
peculiar morphology, and the quantitative structural analysis of these objects indicates
that they are involved in major mergers (Concelice et al., 2005). In addition, the optical
colours, luminosities, sizes and star formation rates derived for SMGs (Blain et al., 2002;
220 Luis Colina
Frayer et al., 2004) indicate that these galaxies are the high-z analogues of the low-z
(U)LIRGs (Arribas et al., 2004). Moreover, typical sub-arcsecond (0.2′′ ) near- and mid-
IR observations of high-z galaxies will correspond to a linear scale of about 1.5 kpc,
similar to the average scale in the low-z sample of ULIRGs for which IFS is available.
Therefore, a good understanding of the physical processes involved in the formation and
evolution of a large fraction of the high-z galaxies will only become possible when a
direct comparison of these galaxies with their low-z counterparts is established both on
the same linear scale and using similar rest-frame wavelength ranges. Initial attempts
at investigating the kinematics of high-z SMGs (Swinbank et al., 2006b) and UV-bright
galaxies (Förster Schreiber et al., 2006) are now being conducted from the ground. Due
to the faintness and small angular size of these high-z systems, these IFS-based studies
are only able to provide preliminary estimates of their dynamical mass and evolution.
6.3 Future IFS facilities: extragalactic studies

Several IFS systems have become operational in different observatories as common
user instruments over the past few years. Most of these systems are installed in large
10 m-class telescopes and are becoming a standard observational technique for almost
all areas of extragalactic research. Future instruments currently under development also
include a second generation of IFS systems, such as the Multi Unit Spectroscopic Explorer
(MUSE; http://www.eso.org/instruments/muse/) and KMOS for the VLTs. These new
instruments will provide new capabilities, combining adaptive optics with either near-IR
multi-object IFS (KMOS) or deep optical IFS over a wide field of view (MUSE). Both
instruments will be ideal for investigating the properties of different populations of galax-
ies in the high-z universe, but also for studying the nuclear regions of low-z galaxies on
scales of few parsecs. IFS systems will also be installed on board the James Webb Space
Telescope (JWST) expanding IFS-based science into the mid-IR wavelength range. The
JWST Near-InfraRed Spectrograph (NIRSpec) and Mid-InfraRed Instrument (MIRI)
will have integral field capabilities covering the near-IR (1–5 µm, NIRSpec) and mid-IR
(5–28 µm, MIRI) range in a low background environment. In the following sections, some
of the relevant characteristics of these future instruments are summarized and some of
the extragalactic and cosmology science that will become available are mentioned.
6.3.1 Optical wide field IFS on the VLT: MUSE

MUSE is a second-generation instrument in development for the VLT of the European
Southern Observatory (Bacon et al., 2004). MUSE, a panoramic IFS, will be operating
in the visible wavelength range from 0.47 to 0.93 µm at a resolution (R) of 2000–4000. Its
main operating mode will be the wide field mode combining IFS over a relatively wide
field of view (1 arcmin2 ) with sub-arcsec spatial resolution (0.3′′ to 0.4′′ FWHM). MUSE
will also have a narrow field mode providing adaptive optics, extremely high angular
resolutions (0.03′′ to 0.05′′ FWHM) over a small field of view (7.5 × 7.5 arcsec2 ).
The main scientific driver of MUSE instrument design is the study of the progenitors
of normal galaxies out to a redshift of 6. These systems are extremely faint and will only
be detected in large numbers through intermediate and deep Lyα line surveys down to a
limiting line flux of 3.9 × 10−19 erg s−1 cm−2 , a factor 100 deeper than currently achieved
with narrow band imaging (for more details, see Bacon et al., 2004). In addition, the
adaptive optics narrow field mode, with a limiting flux of 2.3 × 10−18 erg s−1 cm−2 in
one-hour integrations, will allow us to investigate the parsec-scale kinematics of the
regions surrounding the supermassive black holes in nearby galaxies and the kinematics
of the gas and stellar populations in the central kiloparsec regions of low-z galaxies.
6.3.2 Near-IR multi-object IFS on the VLT: KMOS
KMOS will be a cooled multi-object, near-infrared spectrograph able to position 24
integral field units (IFUs) over a circular field of 7.2 arcmin in diameter. Each of the
IFU will have a 2.8 × 2.8 arcsec2 field of view with a spatial sampling of 0.2′′ and a
spectral resolution (R) of 3500–4000 (http://www.eso.org/instruments/kmos/). Since
bright optical lines such as the [O II] λ3727Å, and the Hα+[N II] complex are shifted into
the near-IR windows for redshifts of up to 5, this instrument will be ideal for investigating
the physical processes that drive the formation and evolution of galaxies, in particular the
mass assembly of galaxies and the connection between active galactic nuclei and galaxy
formation. The predicted sensitivity limits will allow us to map the kinematics of the
ionized gas in high redshift galaxies (z ∼ 1 to 3) brighter than AB ∼ 20 (K) and AB ∼ 22
(J,H) mag arcsec−2 with high signal-to-noise (> 10σ per pixel) in a 10-hour integration
(Sharples et al., 2004).
6.3.3 Near-IR IFS on the JWST: NIRSpec

NIRSpec will be the near-IR spectrograph for the JWST sensitive over the 0.6 µm to
5 µm wavelength range. NIRSpec contains an intermediate spectral resolution (R ∼ 1000
to 3000) IFS, also capable of obtaining low-resolution (R ∼ 100) spectra over the entire
spectral range in one single exposure. This IFS will provide a field-of-view of 3 × 3 arcsec2
with a sampling of 0.1′′ , extending therefore the capabilities of ground-based near-IR
IFS into the 2 µm to 5 µm window. Expected sensitivities are such that emission lines
with fluxes above 2 × 10−21 W m−2 will be detected with a signal-to-noise of 10 in a
ten thousand seconds exposure with a spectral resolution (R) of 2700 (Arribas et al.,
2007). Since the optical Hα emission line shifts into the 2 µm to 5 µm spectral range for
redshifts of up to 6.5, this IFS will be able to investigate the two-dimensional ionization
and kinematic structure of Lyman break galaxies and Luminous Infrared galaxies at very
high redshifts (Arribas et al., 2007; Gardner et al., 2006).
6.3.4 Mid-IR IFS on the JWST: MIRI

MIRI will be the only instrument with IFS capabilities operating in the 5 to 28 µm range
(Wright et al., 2004; European Consortium web page at http://www.roe.ac.uk/ukatc/
consortium/miri/miri.html). In contrast to ground-based mid-IR instruments, MIRI will
be covering this spectral range continuously and will be affected by a thermal background
level orders of magnitude below that of the atmosphere. The entire spectral range cov-
ered by MIRI will be divided in four different channels designed such that the spectral
resolution will be about the same (R of 2400 to 4000). The channels will cover a field of
view of 3.6 × 3.6 arcsec2 to 7.5 × 7.5 arcsec2 with a Nyquist-optimized angular sampling
corresponding to pixels of about 0.2 × 0.2 arcsec2 to about 0.3 × 0.6 arcsec2 . The required
sensitivities guiding the design of the IFS are 1–6 × 10−20 W m−2 for a 10σ emission line
detection in a ten-thousand second exposure (Swinyard et al., 2004). MIRI will therefore
be more than three orders of magnitude more sensitive than the best ground-based mid-
IR instrument on a 10 m-class telescope and about 50 times more sensitive than Spitzer,
while increasing the spatial and spectral resolution by an order of magnitude with respect
to that of Spitzer.
As shown by the IFS-based studies of low-z (U)LIRGs, galaxy interactions can trigger
powerful episodes of star formation and accretion on to nuclear supermassive black holes
that are deeply shrouded in the interstellar medium of the colliding galaxies. These
processes can be studied up to redshifts of 2 to 2.5 through the strong near- and mid-IR
emission lines shifted into the 20 to 28 µm range and to very high redshifts (z > 6.5)
222 Luis Colina
using the Hα+[N II] optical emission lines shifted into the 5 to 7 µm window (Gardner
et al., 2006).
6.4 Concluding remarks

Although still a young technique, IFS is becoming more and more popular and is now used
as a standard tool in a whole variety of research topics covering the low- and high-z Uni-
verse. Compared to pure broad- and narrow-band imaging and to long-slit spectroscopy,
the capabilities of IFS make this technique particularly suitable for the investigation of
the complex two-dimensional stellar, gas and kinematic structure of galaxies. The future
of IFS for extragalactic studies looks very promising. Almost every large 10 m-class tele-
scope is now equipped with at least one IFS, and all large facilities are now designing
future IFS systems including multi-object IFUs and wide-field IFS. Future extremely
large telescopes (diameter of 25 m, or more) will also include IFS systems as part of the
standard instrumentation. There is, however, an important challenge ahead of us before
the power of these systems is fully exploited. These future instruments will produce a
huge amount 10–100,000 of spatially separated spectra per exposure. Therefore, emphasis
has to be made on the development of new sophisticated software tools to automatically
handle the calibration of the raw data and to facilitate their subsequent astrophysical
analysis.
6.5 Acknowledgements
Based on observations with the William Herschel Telescope operated on the island of La
Palma by the ING in the Spanish Observatorio del Roque de los Muchachos of the Insti-
tuto de Astrofı́sica de Canarias. Also based on observations with the NASA/ESA Hubble
Space Telescope, obtained at the Space Telescope Science Institute, which is operated by
the Association of Universities for Research in Astronomy, Inc. under NASA contract
number NAS5-26555.
Much of the work presented here has been supported by the Ministry for Education and
Science of Spain under its National Programme for Astronomy and Space Science. I would
like to thank my collaborators over the years: Santiago Arribas, Ana Monreal-Ibero,
Macarena Garcı́a-Marı́n, and Almudena Alonso-Herrero. Finally, a lot of the material
presented during the lectures was kindly made available by many colleagues who allowed
me to use material given in their PowerPoint presentations. My thanks to Drs Cardiel,
Davies, Falcón-Barroso, Hammer, Jarvis, Kuntschner, Lemoine-Busserole, McDermid,
Mollá, Swinbank, and Villar-Martı́n.
REFERENCES
Arribas, S., Colina, L. (2002), ApJ, 573, 576
Arribas, S., Colina, L. (2003), ApJ, 591, 791
Arribas, S., Mediavilla, E. (1994), ApJ, 437, 149–161
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B. (1996), ApJ, 463, 509
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B., del Burgo, C. (1997), ApJ, 490, 227
Arribas, S., Mediavilla, E., del Burgo, C., Garcı́a-Lorenzo, B. (1999), ApJ, 511, 680
Arribas, S., Colina, L., Borne, K.D. (2000), ApJ, 545, 228
Arribas, S., Colina, L., Clements, D. (2001), ApJ, 560, 160
Arribas, S., Bushouse, H., Lucas, R.A., Colina, L., Borne, K.D. (2004), AJ, 127, 2522
Arribas, S., Ferrint, P., Jakobsen, P. (2007), in Science Perspectives for 3D Spectroscopy.
ESO Astrophysics Symposia, eds. M. Kissler-Patig, M. Roth & J. Walsh. Springer
Bacon, R., Bauer, S.-M., Bower, R. ET AL. (2004), in Ground-based Instrumentation for
Astronomy, eds. A.F.M. Moorwood & I. Masanori, Proceedings of the SPIE, Vol. 5492,
1145
Baldwin, J.A., Phillips, M.M., Terlevich, R. (1981), PASP, 93, 5
Ballo, L., Braito, V., Della Ceca, R. ET AL. (2004), ApJ, 600, 634
Barth, A.J., Shields, J.C. (2000), PASP, 112, 753
Blain, A. W., Smail, I., Ivison, R.J., Kneib, J.-P., Frayer, D.T. (2002), Physics Reports,
369, 111
Bushouse, H.A., Borne, K.D., Colina, L. ET AL. (2002), ApJS, 138, 1
Chapman, S.C., Blain, A.W., Ivison, R.J., Smail, I.R. (2003a), Natur, 422, 695
Chapman, S.C., Windhorst, R., Odewahn, S., Yan, H., Conselice, C. (2003b), ApJ, 599,
92
Charmandaris, V., Uchida, K.I., Weedman, D. ET AL. (2004), ApJS, 154, 142
Cimatti, A., Daddi, E., Renzini, A. ET AL. (2004), Natur, 430, 184
Colina, L., Arribas, S., Borne, K.D. (1999), ApJ, 527, L13
Colina, L., Arribas, S., Borne, K.D., Monreal, A. (2000), ApJ, 533, L9
Colina, L., Borne, K., Bushouse, H. ET AL. (2001), ApJ, 563, 546
Colina, L., Arribas, S., Clements, D. (2004), ApJ, 602, 181
Colina, L., Arribas, S., Monreal-Ibero, A. (2005), ApJ, 621, 725
Conselice, C.J., Bershady, M.A., Dickinson, M., Papovich, C. (2003), AJ, 126, 1183
Conselice, C.J., Blackburne, J.A., Papovich, C. (2005), ApJ, 620, 564
Dasyra, K. M., Tacconi, L.J., Davies, R.I. ET AL. (2006), ApJ, 651, 835
Davies, R. I., Thomas, J., Genzel, R. ET AL. (2006), ApJ, 646, 754
Della Ceca, R., Ballo, L., Tavecchio, F. ET AL. (2002), ApJ, 581, L9
de Zeeuw, P.T., Bureau, M., Emsellem, E. ET AL. (2002), MNRAS, 329, 513
Dı́az-Santos, T., Alonso-Herrero, A., Colina, L., Ryder, S.D., Knapen, J.H. (2007),
ApJ, 661, 149
Dopita, M.A., Sutherland, R.S. (1995), ApJ, 455, 468
Downes, D., Solomon, P.M. (1998), ApJ, 507, 615
Egami, E., Dole, H., Huang, J.-S. ET AL. (2004), ApJS, 154, 130
Emsellem, E., Cappellari, M., Peletier, R.F. ET AL. (2004), MNRAS, 352, 721
Erb, D.K., Shapley, A.E., Steidel, C.C. ET AL. (2003), ApJ, 591, 101
Erb, D. K., Steidel, C.C., Shapley, A.E., Pettini, M., Adelberger, K.L. (2004), ApJ,
612, 122
Farrah, D., Rowan-Robinson, M., Oliver, S. ET AL. (2001), MNRAS, 326, 1333
Ferruit, P., Mundell, C.G., Nagar, N.M. ET AL. (2004), MNRAS, 352, 1180
Flores, H., Hammer, F., Puech, M., Amram, P., Balkowski, C. (2006), A&A, 455, 107
Förster Schreiber, N.M., van Dokkum, P.G., Franx, M. ET AL. (2004), ApJ, 616, 40
Förster Schreiber, N.M., Genzel, R., Lehnert, M.D. ET AL. (2006), ApJ, 645, 1062
Franx, M., Labbé, I., Rudnick, G. ET AL. (2003), ApJ, 587, L79
Frayer, D.T., Armus, L., Scoville, N.Z. ET AL. (2003), AJ, 126, 73
Frayer, D.T., Reddy, N.A., Armus, L. ET AL. (2004), AJ, 127, 728
Gallais, P., Charmandaris, V., Le Floc’h, E. ET AL. (2004), A&A, 414, 845
Ganda, K., Falcón-Barroso, J., Peletier, R.F. ET AL. (2006), MNRAS, 367, 46
Garcı́a-Lorenzo, B., Mediavilla, E., Arribas, S., del Burgo, C. (1997), ApJ, 483, L99
Garcı́a-Lorenzo, B., Mediavilla, E., Arribas, S. (1999), ApJ, 518, 190
Garcı́a-Lorenzo, B., Arribas, S., Mediavilla, E. (2001), A&A, 378, 787
Garcı́a-Lorenzo, B., Sánchez, S.F., Mediavilla, E., González-Serrano, J.I., Chris-
tensen, L. (2005), ApJ, 621, 146
Garcı́a-Marı́n, M., Colina, L., Arribas, S., Alonso-Herrero, A., Mediavilla, E. (2006),
ApJ, 650, 850
Garcı́a-Marı́n, M. ET AL., in preparation
Gardner, J.P., Mather, J.C., Clampin, M. ET AL. (2006), SSRv, 123, 485
Genzel, R., Tacconi, L.J., Rigopoulou, D., Lutz, D., Tecza, M. (2001), ApJ, 563, 527
Giavalisco, M., Steidel, C.C., Macchetto, D. (1996), ApJ, 470, 189
Groves, B.A., Dopita, M.A., Sutherland, R.S. (2004), ApJS, 153, 9
Heckman, T.M., Armus, L., Miley, G.K. (1990), ApJS, 74, 833
Le Floc’h, E., Pérez-González, P.G., Rieke, G.H. ET AL. (2004), ApJS, 154, 170
224 Luis Colina
Lutz, D., Spoon, H.W.W., Rigopoulou, D., Moorwood, A.F.M., Genzel, R. (1998), ApJ,
505, L103
McDermid, R.M., Bacon, R., Kuntschner, H. ET AL. (2006), New Astronomy Review, 49,
521
McDowell, J.C., Clements, D.L., Lamb, S.A. ET AL. (2003), ApJ, 591, 154
Mediavilla, E., Arribas, S., Garcı́a-Lorenzo, B., del Burgo, C. (1997), ApJ, 488, 682
Mediavilla, E., Arribas, S., del Burgo, C. ET AL. (1998), ApJ, 503, L27
Mediavilla, E., Guijarro, A., Castillo-Morales, A. ET AL. (2005a), A&A, 433, 79
Mediavilla, E., Muñoz, J.A., Kochanek, C.S. ET AL. (2005b), ApJ, 619, 749
Monreal-Ibero, A. (2004), Estudio cinemático y de fuentes de ionización de galaxias infrar-
rojas ultraluminosas con espectroscopa bidimensional, PhD, Universidad de La Laguna
Monreal-Ibero, A., Arribas, S., Colina, L. (2006), ApJ, 637, 138
Motta, V., Mediavilla, E., Muñoz, J.A. ET AL. (2002), ApJ, 574, 719
Motta, V., Mediavilla, E., Muñoz, J.A., Falco, E. (2004), ApJ, 613, 86
Mueller Sánchez, F., Davies, R.I., Eisenhauer, F. ET AL. (2006), A&A, 454, 481
Pérez-González, P.G., Rieke, G.H., Egami, E. ET AL. (2005), ApJ, 630, 82
Rieke, G.H., Low, F.J. (1972), ApJ, 176, L95
Sánchez, S.F., Garcı́a-Lorenzo, B., Mediavilla, E., González-Serrano, J.I., Chris-
tensen, L. (2004), ApJ, 615, 156
Sarzi, M., Falcón-Barroso, J., Davies, R.L. ET AL. (2006), MNRAS, 366, 1151
Scoville, N.Z., Yun, M.S., Bryant, P. (1997), ApJ, 484, 702
Sharples, R.M., Bender, R., Lehnert, M.D. ET AL. (2004), in Ground-based Instrumenta-
tion for Astronomy, eds. A.F.M. Moorwood & I. Masanori, Proceedings of the SPIE, 5492,
1179
Soifer, B. T., Sanders, D.B., Madore, B.F. ET AL. (1987), ApJ, 320, 238
Solomon, P.M., Downes, D., Radford, S.J.E., Barrett, J.W. (1997), ApJ, 478, 144
Swinbank, A.M., Smith, J., Bower, R.G. ET AL. (2003), ApJ, 598, 162
Swinbank, A.M., Smail, I., Bower, R.G. ET AL. (2005), MNRAS, 359, 401
Swinbank, A.M., Bower, R.G., Smith, G.P. ET AL. (2006a), MNRAS, 368, 1631
Swinbank, A.M., Chapman, S.C., Smail, I. ET AL. (2006b), MNRAS, 371, 465
Swinyard, B.M., Rieke, G.H., Ressler, M. ET AL. (2004), in Optical, Infrared and Millimeter
Space Telescopes, eds. J.C. Mather, Proceedings of the SPIE, Vol. 5487, 785
Tacconi, L.J., Genzel, R., Lutz, D. ET AL. (2002), ApJ, 580, 73
Tecza, M., Baker, A.J., Davies, R.I. ET AL. (2004), ApJ, 605, L109
Tully, R.B., Fisher, J.R. (1977), A&A, 54, 661
Urry, M., Padovani, P. (2000), PASP, 112, 1516
van Dokkum, P.G., Franx, M., Förster Schreiber, N.M. ET AL. (2004), ApJ, 611, 703
Veilleux, S., Kim, D.-C., Sanders, D.B. (1992), ApJS, 522, 113
Villar-Martı́n, M., Sánchez, S.F., De Breuck, C. ET AL. (2006), MNRAS, 366, L1
Wright, G.S., Rieke, G.H., Colina, L. ET AL. (2004), in Optical, Infrared and Millimeter
Space Telescopes, eds. J.C. Mather, Proceedings of the SPIE, 5487, 653
7. Tutorials: How to handle 3D
spectroscopy data
SEBASTIÁN F. SANCHEZ, BEGOÑA GARCÍA-LORENZO AND
ARLETTE PÉCONTAL-ROUSSET
7.1 Introduction
Integral field spectroscopy (IFS) is a technique to obtain both spatial (x,y) and spec-
tral (λ) information of a more or less continuous area of the sky simultaneously on the
detector. Only a few instrumental concepts allow 3D information on 2D detectors to be
obtained, and all of these are based on field splitters such as fibre bundles, lens array,
or image slicers (see Figure 7.1) to sample the field of view. Each sampled element is
then dispersed using a classic spectrograph and produces a spectrum on the detector.
Depending on the field splitter used, the geometry of the spectra on the detectors may be
very different. This diversity leads to the creation of very specific reduction techniques
and/or packages, i.e. one per instrument built (e.g. P3d, Becker, 2001). Combined with
the inherent complexity of 3D techniques, such software diversity has reduced the use of
IFS for decades to a handful of specialists, mainly those involved in the teams building
such instruments.
Conscious that this would be a handicap IFS specialists Walsh and Roth (2002) have
started to standardize techniques and tools for integral field units (IFU). Recently, the
Euro3D Research Training Network (RTN), whose aim was to promote 3D spectroscopy
all over Europe (Walsh and Roth, 2002), made a great effort to create a standard data for-
mat (Kissler-Patig et al., 2004) for storing and exchanging 3D data, developing an applica-
tion programming interface, API (Pécontal-Rousset et al., 2004), to ease the use of such a
data format and creating a visualization tool (Sánchez, 2004) usable by any existing IFU.
All these tools are freely distributed to the community (http://www.aip.de/Euro3D).
However, there is still a lack of software packages of general use for most of the data
reduction and data analysis processes (reduction steps, line fitting, gas and stellar
kinematics, etc.) of 3D data. The astronomical community faces a plethora of differ-
ent pipelines and tools specific to each instrument. We have developed a set of tools and
procedures useful for most of the IFS data; these will be explained in these tutorials,
as they were presented in the Instituto de Astrofı́sica de Canarias (IAC) Winter School.
Many of these tools are still in the development process, and we are distributing the
current version without any guarantee that they will work in the optimal way. The aim
is to provide simple and standardized tools, and also to teach the user the main logical
steps he or she will face in trying to reduce and then analyse his or her 3D data.
! Section 7.2 describes the most widely used data formats to store 3D spectroscopy
The layout of the chapter will be as follows:
! Section 7.3 describes some of the existing tools to analyse 3D data before the Euro3D
data.
! Section 7.4 describes the installation requirements of the different packages used in
network.
! Section 7.5 introduces E3D, the Euro3D visualization tool (Sánchez, 2004).
these tutorials.
! The sky background subtraction is described in Section 7.6.

! Section 7.7 introduces some notions of emission-line fitting and gas kinematics.
! Crowded field spectroscopy and 3D modelling is described in Section 7.8.
225
226 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
Figure 7.1. Layout of the main techniques used to perform integral field spectroscopy
(Allington-Smith and Content, 1998).
! Section 7.9 describes the empirical correction of the differential atmospheric
! Section 7.10 describes the data reduction process with fibre-feed spectrographs.
refraction.
7.2 Data formats for storing IFS reduced data

IFS data can be stored using different data formats, as far as it is possible to store the
spectral information associated with the position on the sky, to which it belongs, and the
shape of the spaxel (which stands for sampled spatial element; see Kissler-Patig et al.,
2004). The various existing data formats were designed by software developers to meet
instrument specifications, and are, in general, only valid for a specific instrument. The
exchange and comparison of data from different integral field instruments was therefore
difficult to achieve.
After several discussions and brainstorming sessions, the Euro3D network created a
standard format, the Euro3D format (Kissler-Patig et al., 2004), whose aim is to share
data and analysis tools. This data format is based on FITS (Flexible Image Transport
System), a widely used data format in the astronomical community all over the world.
It consists mainly in a binary table, where the position of the spatial sampled elements
(spaxels) on the sky are stored on the same row as the associated spectra. Although
this format is now intended to be a standard for any reduced 3D data, other formats
(pre-Euro3D network) are still in use. Two of them are widely used for IFS: datacubes
(3D image) and row-stacked spectra (RSS hereafter).
Datacubes are only valid to store reduced data from instruments that sample the sky
plane in a regular grid or for interpolated data. The data are stored in a 3D FITS image,
were the x and y axes stand for the two spatial dimensions, and the Z axis corresponds
to the dispersion axis.
RRS format is a 2D FITS image where the x and y axes contain the spectral and
spatial information regardless of their position in the sky. It requires an additional file
Tutorials: How to handle 3D spectroscopy data 227
(a FITS or ASCII table), where the position of the different spatial elements on the sky
is stored. RRS is widely used by IFUs where spectra are well-separated on the detector
(e.g. INTEGRAL, GMOS or Gemini Multi-Object Spectrograph).
7.3 Tools before the Euro3D network

The analysis of 3D data is not an easy task due to the huge amount of information to
process. Before the Euro3D network, there was no dedicated software to analyse 3D data,
in part due to the lack of a standard format for storing IFU data as we have already men-
tioned. Scientists working with 3D used to combine well-known and widely used spread
astronomical software like IRAF (Image Reduction and Analysis Facility), MIDAS (Mid-
Infrared Asteroid Spectroscopy), STARLINK, etc., with home-made software developed
using their favourite programming language (fortran, C, IDL, etc.). A common procedure
was, for example, to have a look to 3D data using IRAF (with the DS9 visualization tool),
then fit spectral lines using DIPSO, and create images with a home-made FORTRAN
program.
In this section, we will briefly describe some ways to visualize and analyse 3D data
using traditional astronomical packages, but many other software packages may also be
used. The user should always keep in mind that these tools were not specifically designed
to handle IFS. However, some of them have recently included a few tasks to process 3D
data due to the increasing popularity of 3D techniques in the last years.
7.3.1 DS9 visualization tool

DS9 (http://hea-www.harvard.edu/RD/ds9/) is an astronomical imaging and data visu-
alization application. It handles FITS images and binary tables, multiple frame buffers,
region manipulation, many scaling algorithms and colour maps, mosaic images, tiling,
blinking, geometric markers, arbitrary zoom, rotation, pan, and a variety of coordinate
! To display an RSS image with DS9, first select the File menu and then select Open.
systems.
! To display datacubes (i.e. 3D images) with DS9, do the same as above (i.e. select the
File menu and then the Open option). Once you have choosen the file to be loaded,
a pop-up window comes up entitled ‘Data Cube’, with a scroll bar that allows you to
scroll through spectral pixels (i.e. wavelength). Moving that scroll bar, DS9 displays
! To display datacube stores in multi extension FITS file, use the File menu and then
the slice corresponding to the selected wavelength (monochromatic image).
the ‘Open Other’ option and the ‘Open Multi Ex as Data Cube’ option.
DS9 can be also run from the shell, entering ds9 followed by the name of the FITS
image (e.g. ds9 test.fits). For multiple extension FITS files, use the metacube option of
DS9 (e.g. ds9 -medatacube test.fits).
7.3.2 Checking 3D data using IRAF

IRAF (http://iraf.noao.edu/) is a popular tool to reduce, visualize and analyse astronom-
ical data. IRAF needs a visualization tool such as DS9 (Section 7.3.1) to display data.
It is possible to visualize datacubes slice by slice through the IRAF command window
using the display command. The user should know the structure of its data FITS file.
Let’s say the datacube contains more than two axes (i.e. NAXES>2 in the FITS header).
Use
display test.cube.fits[*,*,n],
Figure 7.2. DS9 displaying an RSS image. The different rows correspond to the various spectra
of a particular area of an object (an active galaxy in this case). The horizontal axis is wavelength.
Vertical lines covering all the spectra are sky lines. A few emission lines from the observed object
can also be identified in this frame.
n being the value of the third dimension to display. If it is a multiple extension format,
use
display test.cube1.fits[n][*,*],
where n is the extension number to display.
You can also inspect the different spectra using the splot command:
splot test.cube.fits[10,10,*].
The RSS images (2D images) may also be visualized using the display command:
display test.ms.fits.
Figure 7.3. Visualizing 3D data using IRAF and imarec. The snapshot of the console shows
the different IRAF tasks used to visualize RSS data.
To have a look at the spectra, use the command splot:

splot test.ms.fits
and select a spectrum. It is possible to overplot other spectra using the options of the
splot command.
However, IRAF has no routine to visualize monochromatic (data from a given wave-
length) or filtered images from RSS data. The INTEGRAL team1 has developed
a task called imarec (which stands for image reconstruction) to extract and view
monochromatic images from RSS images and ASCII files containing the positions of
the spectra. Imarec is based on standard IRAF routines and a FORTRAN program
(more information on IRAF routines developed for INTEGRAL data can be found at
http://www.iac.es/proyect/integral/). Imarec prompts the user for some parameters and
produces a filtered image that can be displayed in DS9. Figure 7.3 shows the different
steps to visualize RSS data in IRAF.
1
INTEGRAL (http://www.iac.es/proyect/integral/) is a 3D instrument at the William
Herschel Telescope (WHT) (Arribas et al., 1998).
7.3.3 Starlink
The STARLINK project (http://starlink.rl.ac.uk/) compiled (until 2005) a set of interac-
tive data reduction and analysis tools for astronomical data. For example, in the STAR-
LINK software package list we can find FIGARO, a great reduction package for reducing
spectroscopic data, which also includes image and datacube manipulation capabilities
(for more information see http://starlink.rl.ac.uk/star/docs/sun86.htx/sun86.html). By
default, FIGARO accesses data files in STARLINK’s NDF (Extensible N-Dimensional
Data Format) data format, but it can also access other data formats, like FIGARO’s old
DST (Data Storage Technology) format, FITS and IRAF formats. By the way, IRAF is
a software facility of the STARLINK project. Check the following web page for a com-
plete list of STARLINK software: http://starlink.rl.ac.uk/static www/soft2 LSF.html.
Unfortunately, the STARLINK project has already finished and software packages such
as FIGARO are unsupported.
As an example of the STARLINK software capabilities for analysing 3D data, we will
use DIPSO, a friendly spectrum analysis program, for fitting lines. Once the user has
started the STARLINK and DIPSO sessions, the plotting device should be defined using
the command dev xwindow. Then the alasrd command reads an ASCII file containing
the spectral data and relevant information. The initial values of the fitting parameters
(e.g. central wavelengths, full-width-at-half-maximum or FWHM, etc. of the different
line components) are defined using the task elfinp. Finally, the elfopt command runs the
fit. Figure 7.4 shows both the complete sequence of DIPSO commands for fitting lines on
a spectrum and a plot of the results. Unfortunately, for 3D data this procedure should
be done once per spectrum and the number of spectra may be huge. Moreover, after the
fitting, maps of the line parameters (such as intensity, and FWHM) are highly desirable,
but neither DIPSO nor any other STARLINK tasks can do these. Other software tools
must be used to create such maps.
7.3.4 Available software for specific instruments

Currently, a large number of integral-field spectrographs (IFSs) are operated around the
world and, most of the time, the team who built the instrument provides the community
with dedicated data reduction software; however, few of them provide tools to analyse
the data.
One example is the set of XOasis/XSauron and XSnifs software (see, for example,
http://www.ing.iac.es/Astronomy/instruments/oasis/xoasis.html), the data reduction
and processing packages designed to handle data from the OASIS (Optically Adaptive
System for Imaging Spectroscopy), SAURON (Spectrographic Areal Unit for Research
on Optical Nebulae) and SNIFS (Supernovae Integral Field Spectrograph) instruments
built at the Centre de Recherche Astrophysique de Lyon (CRAL). OASIS (http://www-
obs.univ-lyon1.fr/labo/oasis) is a multi-mode spectro-imager, previously operated on
the CFHT, and now installed at the WHT. SAURON (http://www.strw.leidenuniv.nl/
sauron) is another IFU installed at the WHT and optimized for studies of the stellar
kinematics, gas kinematics and line-strength distributions of nearby spheroids. The last
one, SNIFS; http://snfactory.lbl.gov, is currently operated at the University of Hawaii
(UH) telescope at Mauna Kea and dedicated to the survey of Type Ia supernovae (part
of the SNFactory project, whose aim is to measure the expansion history of the Universe
and explore the nature of Dark Energy).
All of these software codes share the same development basis (using C, Fortran and
tcl/tk), and were designed to be run through a pipeline, or via shell scripts, or in interac-
tive mode using a more user-friendly interface (see Figure 7.5). Through a menu-driven
Figure 7.4. Line fitting using DIPSO (from the STARLINK software).
graphical user interface (GUI), they provide the user with all the necessary data reduction
steps, plus a set of generic graphical tools to view or plot data, assign keywords, etc. They
also provide analysis tools like image reconstruction (maps), spectra summation over an
aperture, line fitting (developed externally and currently being incorporated), merging
of datacubes, etc. They all handle different data formats, and are currently debugged to
fully support Euro3D. Binaries of the XOasis software are made available to the user com-
munity and may be downloaded from the Oasis Webserver, while others may be obtained
through the consortia who develop the instruments. The last ones usually contain
more sophisticated analysis tools like line-of-sight velocity distribution (LOSVD) fitting,
etc.
Another example is IDA. INTEGRAL team developed IDA, the Integral Data Anal-
ysis tool (Garcı́a-Lorenzo et al., 2002). IDA was written in interactive data language
(IDL) and offers a widget-based interface. IDA uses an interactive display tool, ATV
(for more information about ATV, see http://www.physics.uci.edu/∼barth/atv/), which
allows control of images similar to other display tools like DS9. IDA plots individual
spectra, creates and displays maps, overplots spectra on recovered images, zooming in
and zooming out, and calculates velocity fields through the cross-correlation technique.
Although IDA was written to work with INTEGRAL data, it can be also used for similar
data from other IFSs through minor software changes. Figure 7.6 shows some of the IDA
capabilities for analysing INTEGRAL data.
Figure 7.5. XOasis Graphical User Interface.
7.3.5 Exercises
In this section, we present a few exercises to visualize 3D data using traditional astro-
! RSS → NAME RSS.fits (with associated position table NAME RSS.pt.txt);

nomical software. Hereafter, we use the following filename convention:
! datacube → NAME.cube.fits;
! Euro3D format data → NAME.e3d;
! Position Table → NAME.pt.
Using IRAF
To run IRAF, type cl or ncl at any xterm or xgterm:
cl> imred NAME RSS.fits
cl> specred NAME RSS.fits
It is possible to zoom in, zoom out, fit lines, plot different lines, with all the options of
the gtools. For more help, see help/splot.txt and help/gtools.txt
Using DS9 through IRAF

To display datacubes slices to slices.
cl> disp NAME.cube.fits[*,*,500]
To display RSS spectra:
cl> disp NAME rss.fits
Figure 7.6. The INTEGRAL Data Analysis (IDA) tool capabilities for visualizing and
analysing 3D data.
To display spectra from a datacube.

cl> disp NAME.cube.fits[10,*,*]
cl> splot NAME.cube.fits[10,10,*]
Using DS9 outside of IRAF (datacubes)

To display datacubes, using DS9 GUI:
Menu: File → Open
Once the file is loaded, a pop-up window comes up, with a scroll bar allowing the
user to scroll over the spectral pixels (i.e. wavelength). While moving the scroll bar, DS9
shows the slice corresponding to the selected wavelength.
7.4 Installation of the E3D software

The tutorial will use various software tools, some of them being part of the Euro3D
network activities. The software (as it was distributed at the Winter School) can be
downloaded from: http://www.iac.es/winters/2005/software3D.html. In summary, the
! E3D, the Euro3D visualization tool (Sánchez, 2004);

software tools distributed are:
! FIT3D, a fitting tool for 3D spectroscopy (Sánchez et al., 2004; Sánchez et al., in
! R3D, a software package for reducing fibre-based IFS (Sánchez et al., in preparation).
preparation);
They all have a C-core written using the I/O Lyon-C Library and its development envi-
ronment (Pécontal-Rousset et al., 2004). The PGPLOT graphical library (Pearson, 1995)
is used also for plotting purposes and, in some cases, Tcl/Tk libraries for GUIs. In many
cases, perl is also used as a scripting language, or as a prototyping environment.
! gcc, the GNU C compiler

Here is the list of libraries and binaries required to build or run such tools:
! g77, the GNU Fortran77 compiler2

! Tcl/Tk, version 8.3 at least. http://www.tcl.tk
! PGPLOT, (http://www.astro.caltech.edu/∼tjp/pgplot), installed with the fol-
lowing drivers: XSERVER, XWINDOWS, POSTSCRIPT, GIF and XTK. The required
! LCL I/O library, version 1.2, http://www.aip.de/Euro3D

drivers.list is distributed with the software.
! Astro::FITS::CFITSIO.
and the following perl modules (http://www.cpan.org).
! Math::Approx.
! Math::Derivative.
! Math::FFT.
! Math::Matrix.
! Math::Spline.
! Math::Stat.
! PDL.
! PGPLOT.
! Statistics::OLS.
We recommend the user installs the source code of all these packages. You can down-
load this from the following web page: http://www.iac.es/winters/2005/software3D.html.
When available, pre-compiled binaries are also included for both Linux or Mac OSX, to
ease the installation for users who are not very familiar with software building, and who
are using standard Linux/Mac OSX distributions.
7.4.1 Installing the LCL library

The Lyon-C library (LCL; Pécontal-Rousset et al., 2004) is a C-coded I/O library that
handles multi-extensions fits files in many ways (1D, 2D and 3D images, ASCII or
binary tables and the Euro3D format). It also includes a simple and useful command-line
argument-handling system. It is distributed through the Euro3D web page (http://www.
2
The new gcc compiler does not include anymore g77, so you must install it separately.
aip.de/Euro3D/). A copy of version 1.2 is included there (E3D io LCL-1.2.tar.gz). To
built it, just follow these instructions:
(i) First create a directory in your computer, that we name hereafter as yourDIR. Copy
the compressed tar file into this directory yourDIR.
(ii) Uncompress and untar the file.
(iii) Go to the created directory, e.g. cd E3D io LCL-1.2
(iv) Run the configure program: ./configure
(v) Run the make program: make
(vi) Define the environment variables: IFU PATH and IFU DEFAULT FMT; e.g. for a tcsh
shell:
setenv IFU PATH yourDIR3 ;
setenv IFU DEFAULT FMT Euro3D;
(vii) To check the library, go to the directory:
cd checklib/exec
and run the test program:
./check E3D io
If you do not have any error messages, then the install is OK.
7.4.2 Installing PGPLOT

PGPLOT is a fortran graphic library (Pearson, 1995) that has a C and a perl wrap-
per, which allow the user to use it also when coding with these languages. From
http://www.astro.caltech.edu/∼tjp/pgplot/ you may download the latest version of
PGPLOT. The instructions on how to install it are described in the INSTALL.txt file
included in the tarball.
7.4.3 Installing E3D

E3D (Sánchez, 2004) is a visualization tool for IFS created through the Euro3D
RTN. We distribute in these proceedings the lastest version, v1.3b: http://www.iac.es/
winters/2005/software3D.html in two formats: the source code (e3d-1.3b.tar.gz),
and a binary distribution, for both Linux (E3D v1.3b linux.tgz) and Mac OSX
(E3D MacOs.tgz). Binary distributions include most of the required libraries, although
some of them have to be installed.
Installing E3D from the source code

To install, follow the instructions below:
(i) Copy the tar file into the Euro3D directory (which is referred to using the previous
environment variable $IFU PATH).
(ii) Using the command tar xvfz e3d-1.3b.tar.gz, extract the archive file
e3d-1.3b.tar.gz under $IFU PATH. It will create the directory e3d-1.3b (or a
similar one depending on the distribution number you are installing).
(iii) Change directory to the newly created one (cd e3d-1.3b), and execute the configure
and make scripts:
./configure
make
3
The IFU PATH should refer to yourDIR and not to yourDIR/E3D io LCL-1.2.
(iv) Go to the scripts directory (i.e. cd scripts) and execute the install script
(./install.pl). This script will customize the tk e3d.tcl according to your direc-
tory tree.
(v) Add the path to the directory user/bin (under the E3D directory) to your environ-
ment variable PATH.
(vi) Test your E3D installation by executing the script tk e3d.tcl, and then loading
the file test.e3d that you will find under the scripts directory.
Installing the binary distributions of E3D

Follow the instructions below:
(i) Copy the tar file into the Euro3D directory (which is referred to using the previous
environment variable $IFU PATH).
(ii) Uncompress the tarfile.
(iii) Go to the newly created directory.
(iv) Execute the ./install static.pl, and follow the instructions.
(v) Test your E3D installation by executing the script tk e3d.tcl, and then loading
the file test.e3d that you will find under the scripts directory.
7.4.4 Installing FIT3D

FIT3D is a fitting tool (Sánchez et al., 2004; Garcı́a-Lorenzo et al., 2005) to fit emission
lines and continuum spectrum, in order to analyse gaseous emission and stellar kinemat-
ics. It consists of a core coded in C for the low-level fitting routines, and a wrapper written
in perl to facilitate the creation of the configuration files, describing the model to fit.
(This is an idea it shares with DIPSO, to create the fitting model using a configuration
file.) Several emission lines and continuum backgrounds can be fitted simultaneously,
with different kinematics, setting bounds to the parameters of the emission lines, linking
parameters between lines for creating kinematics systems (i.e. several lines sharing the
same kinematics properties), and deblending different emission line components. It uses
datacubes and row-stacked spectra as input data, and produces FITS files as output. In
some cases, it requires E3D for the creation of the interpolated maps.
Based on the LCL development environment, it shares the same requirements with
E3D. To install it, just use the same installation instructions as above. At the website:
http://www.iac.es/winters/2005/software3D.html we distribute two versions of FIT3D:
the source code (fitting tool-1.0.tar.gz), and the binaries distributions for Linux
(FIT3D v1.0 linux.tgz) and Mac OSX (FIT3D v1.0 MacOsX.tgz). Except for the C-
core, FIT3D uses perl routines that require the modules listed above to be installed (see
Section 7.4).
7.4.5 Installing R3D

R3D is a package for reducing IFS data obtained with any fibre-based IFU (Sánchez,
2006). Although valuable efforts were made by the Euro3D network to produce stan-
dardized tools to analyse data, there is still a lack of generic packages for data reduction
(which was beyond the scope of the network). The astronomical community encounters a
wide variety of pipelines specific to given instruments, whose procedures and parameters
differ strongly, and whose outputs are difficult to compare. Following preceding efforts
(e.g. P3d, Becker, 2001), we have developed R3D, a reduction package able to reduce any
fibre-feed IFU data.
We first coded the algorithms in Perl, using the Perl data language (PDL, Glazebrook
et al., 2001), in order to speed up the algorithm testing phase. Once the algorithms
were tested with both simulated and real data, we recoded the most time-consuming
routines in C to improve their performance. The Perl version of R3D is fast enough to
produce valuable science frames in a reasonable elapsed time, although for instruments
with many fibres (like VIMOS or Visible Imaging Multi-Object Spectrograph), or for
some slow processes like cross-talk correction algorithms, we strongly recommend the
C-version. A major advantage of using Perl and C is that they are freely distributed and
platform independent, and therefore R3D can be installed under almost any architecture
without major efforts and at no cost. The instructions on how to download and install R3D
can be found at http://www.caha.es/sanchez/r3d/, where we keep the different versions
updated.
Installing the perl version of R3D

R3D uses perl routines that require the modules listed above to be installed (see Section
7.4). We distribute a tar file including the routines and some documentation (R3D.tgz).
To install them, uncompress the tar file and follow the instructions included in the
INSTALL.txt file.
Installing the C version of R3D

Based also on the LCL development environment, it has the same requirements as
E3D. To install it, just follow the same installation instructions. We distribute two
versions of R3D: the source code (r3d-0.1.tar.gz) and a Linux binary distribution
(R3D c linux-0.1.tgz). Except for the C-core, FIT3D uses perl routines that require
the modules listed at the beginning of Section 7.4.
7.5 E3D: the Euro3D visualization tool

E3D is a generic visualization tool for IFS data, fully independent of the instrument used
to obtain such data. It was created on behalf of the Euro3D network (Sánchez, 2004),
and thus handles the Euro3D data format as default format but is able to work with
others formats too (like datacubes and RSS data).
! The Euro3D.o library is a type of application programming interface (API) of low-

E3D has a C-coded core, and may be divided in three logical blocks:
level functions, including routines calling the Euro3D I/O library, routines to access
the shared memory server (SHM), and basic functions needed to plot and analyse
the data. These functions may be invoked by any external program, which needs to
! The tk e3d Tcl/Tk interpreter. This is a stand-alone program embedding the Tcl/Tk
read/write Euro3D data, or uses the SHM.
interpreter with Euro3D routines. These routines wrap the previous C-functions in
Tcl, and can be used, within the tcl interpreter, to load/save Euro3D files, plot
single or co-added spectra, plot monochromatic/polychromatic maps, interpolate
! A number of stand-alone tools, which may be used from the command line, may be
these, save maps in FITS format, etc.
helpful in accessing Euro3D files. Perhaps the most interesting routine is any2Euro3D,
converting single-group IFS data from any instrument into the Euro3D Format.
Together with this C-core, we provide a Tcl/Tk GUI, which calls the Euro3D-Tcl routines
(tk e3d.tcl). The GUI is organized around three different windows corresponding to
! The Stacked Spectra Inspector, or main window, includes the main menu with various
three different ways of investigating the data:
options to handle the input and output data and to pop up the other windows. Once
loaded, the data are automatically plotted in its canvas, using the stacked-spectra
representation. Moreover, a command-line prompt has been included in this window
Figure 7.7. E3D Stacked Spectra Inspector: this is the E3D main GUI window, from which
all the others can be created. It displays all the spectra contained in the 3D file stacked row by
row.
Figure 7.8. E3D Spaxels Inspector: this is the GUI provided for plotting monochromatic or
polychromatic maps. It is possible to select different spaxels, to be displayed subsequently on
the Spectral Inspector.
to directly call any Euro3D-‘tcl flavor’ routines. Figure 7.7 gives a snapshot of this
! The Spaxels Inspector provides the user with a main canvas for viewing
window.
monochromatic/polychromatic datacube slices or maps, which were selected from

the main window or from the Spectral Inspector (see below). It includes eight minor
canvasers for buffering the most recently created maps. It also includes a menu to
handle spaxels/maps options, including spaxels selection, object creation and differ-
ent methods to view the data. Figure 7.8 shows a screen shot of such a window.
Figure 7.9. E3D Spectral Inspector: this is the GUI provided for plotting spectra.
! The Spectral Inspector has a canvas for viewing/plotting the spectra corresponding
to the spaxels selected from any of the two previous windows. It also includes a
menu with different options to handle the spectra. Figure 7.9 gives a snapshot of
this window.
7.5.1 Working with E3D: the GUI
We describe here, with a few examples, the capabilities of E3D. As we quoted above, E3D
can handle data in the Euro3D format, its default format, or use the RSS and datacube
formats. All of these are converted internally to the same structures, so, once loaded,
! Loading data in the Euro3D format:

they are all handled in the same way.
– In the top menu of the main window, select Load in the File menu. A pop-up
window appears.
– Change from default fits extension to e3d.
! Loading data in the RSS format:

– Select the file NAME.e3d.
– In the top menu of the main window select Import RSS in the File menu. A
pop-up window appears.
– Select the NAME RSS.fits RSS file.
! Loading datacubes:
– Select the NAME RSS.pt position table.
– In the top menu of the main window select Import cube in the File menu. A
– Select the file NAME.cube.fits.
You can now follow the exercises. In order to be able to run the exercises, first change to
the directory tut1 and open the E3D GUI running the tk e3d.tcl command.
Exercise 1: Try to load your own data. First identify what format your data are
in (read the reduction tool notes if necessary).
If you need to build a position table, you have to know that the position tables in E3D
are ASCII files with four columns, where the first row gives the geometry of the dataset
and the remaining rows contain the positions, as below:
TYPE SIZE1 SIZE2 TYPE ID
INDEXi Xi Yi TYPE IDi
...
TYPE is the shape of the spaxels and can be: R for a rectangle shape, S for a square, C
for a circle, and H for a hexagon. Here is an example of such a table:
H 0.2 0.2 1
1 0.0 0.0 1
2 0.5 0.0 1
3 0.5 0.5 1
...
7.5.2 Quick look over the various windows of the E3D GUI
As quoted before, the GUI of E3D has three windows corresponding to various inspections
of the data:
The main window or stacked-spectra inspector

This is the window which popped up first when running E3D, and where the data are
displayed as stacked spectra (RSS format), using the same order in which they are stored
on disk (i.e. no preferential spatial direction, a priori). Using this window, it is possible
to browse the spectra.
! Change the directory to tut3 and run E3D.

Perform the following instructions:
! In the top-level menu of the main window, select Import cube in the File menu. A
! Select the file NAME.cube.fits.

By default, the main window displays only 256 spectra (i.e. not the total number of
spectra) as a 2D image, where each row represents a single spectrum. E3D automatically
sets the cuts of intensities for the look-up table, determining the median and the standard
deviation of the intensities plotted. Sometimes, the estimate is not good enough and it is
better to set the values manually. To change these maximum and minimum values, just
fill the Max and Min entries in the left-hand menu. Note: A range of [−1200] would be
acceptable to visualize the data. For a better contrast it is sometimes useful to tune the
Brightness and Contrast settings too, edit the corresponding entries, or just move the
corresponding scroll bars.
Exercise 2: Change the maximum, minimum, brightness and contrast settings until you
get a proper visualization of your data.
To browse the spectra (in cases where you have more than 256 spectra) you can move the
scroll bar in the bottom part of the left-hand menu, below the label Range of Spectra
plotted. It is also possible to change the number of plotted spectra, editing the values
of the two entries labelled First (the first spectrum plotted) and Range (the number of
spectra plotted).
Exercise 3: Let’s play with this browsing capability.
It is possible to zoom over a given spectral region, by selecting this region and clicking
twice over the main display with a left click of the mouse. The selected wavelength range
appears in the W1 and W2 entries in the left-hand menu. You can also edit these entries
directly, if you want a more accurate selection of the wavelength bounds. In order to
return to the default range click on the Release button.
E3D always uses the same method to select elements. The left button of the mouse is
used to select a window and the right button for sending this selection to another window
(inspector). We will experiment with this later on.
Exercise 4: Play with the zooming capabilities until you get used of them.
The combination of zooming capability and range selection is useful to look at specific
wavelength and spatial ranges; e.g. for emission line objects.
Exercise 5: Try to locate some of the emission lines in this object.
The spaxels inspector

The spaxels inspector is the window used to visualize the spatial distribution of the inten-
sities for each spaxel, at a selected wavelength range. It allows you to select individual
(or groups of) spaxels/channels to view the spectrum/spectra corresponding to them.
On the main window menu, click on the Open option of the Spaxels submenu. A new
window pops up: this is the spaxels inspector. It consists of a top menu bar, a left-hand
menu, with similar entries to the one of main window, and then other entries specific to
this inspector. There is also a main display, to display slices (or maps), and eight small
buffer windows, where the last eight displayed slices are stored. Row.No indicates the row
number in the RSS representation of the spaxel over which the cursor is located. Flux,
X and Y show the corresponding flux, and the coordinates of the spaxel. The flux is the
average value of the intensities of the spectra for the selected wavelength range.
As a general remark, E3D displays average values when a multiple selection (several
spectral channels or spaxels) is done. This feature has the advantage that, for subtracting
wavelength ranges, it is not necessary to know the width of the subtracted range. It is
however good to note this width somewhere, when saving the data, because it may be
useful for future computations. N.SPAXELS is the number of currently selected spaxels
and MEAN and SIGMA show the mean flux value and its corresponding standard deviation
over the selected channels. We will explain below how to select spaxels.
To display a slice, go back to the main window (by clicking on it) and select a wave-
length range. Once selected, right-click on the mouse. The reconstructed slice, computed
by averaging the intensities over all the spectra at the selected wavelength range, is
displayed in the spaxels inspector.
Exercise 6: Select various wavelength ranges, and look at the differences between the
slices.
One of the strange (and useful) features of the spaxel inspector is the compatibility
of selecting a narrowband image and creating a movie effect by clicking on the display
on the main window. Keeping the central mouse button pressed, move the mouse along
the spectral direction (browsing the wavelengths). Once the button is released, the most
recently selected range is stored in the spaxels inspector.
Exercise 7: Play with the movie effect to see the the object shifted from the blue to the
red due to the differential atmospheric refraction (DAR) effect, and the areas dominated
by the emission lines.
It is possible to go back to one of the previous displays (from the last eight displayed at
least) by right-clicking on the mouse on the corresponding buffer (small display), at the
bottom of the spaxels inspector window.
To select a spaxel (or a group of them), just left-click on the mouse and keep it pressed
while moving over the spaxel you want to select. A hashed polygon will appear over the
selected spaxels. Note that the N.SPAXELS, MEAN and SIGMA entries change when selecting
the spaxels. More selection options will be described below. To clear a given selection,
just click on the Clear option of the Select submenu in the Spaxels Inspector top menu
bar.
Exercise 8: Play with the spaxels selection capability.
The spectra inspector

The spectra inspector is used to display and plot selected spectra (possibly in a different
order to the ones that are stored on disk). On the main window menu, click on the Open
option of the Spectra submenu. The new window that pops up is the spectra inspector.
It consists of a top menu bar, a main display and a left-hand menu. Most of the entries
in the left-hand menu are similar to those of the main window, except for the spectra
box list, which will be explained below.
The easy way to select a group of spectra is to come back to the spaxels inspector
and make a selection of spaxels, as described above, then to right-click on the display
of the spaxels inspector. The spectra corresponding to the selected spaxels will be
displayed. By default, the spectra inspector shows, on the upper part of the display, a
stacked-spectra image of the selected spectra in the sequential order used for the selection
and, on the bottom, a plot of the average of these spectra.
Exercise 9: Practise using the spectra inspector. Change the maximum and minimum,
the brightness and the contrast in order to visualize the selected spectra properly.
As in the main window, it is possible to zoom over spectral regions and/or select them,
using the same settings as above. You can also display the average slice of a selected
number of spectra by right-clicking on the mouse.
7.5.3 Main-menu options

The top menu-bar of the main window presents the various options that affect the dataset
(except for the sky subtraction, which is part of the spectra inspector). It also shows
the setup of the visualization parameters.
The File menu

This contains all the features to input or output the data on disk. The Euro3D data
format (Kissler-Patig et al., 2004) is the native and default format of this visualization
tool, so the load/save options are intended for use with that format. Other data formats,
like the RSS and datacube formats, can be imported/exported using the import/export
menu option.
When exporting data to a datacube, E3D always performs an interpolation using the
method currently set in the main window (configuration). If you import a datacube and
export it again, it does not keep the previous format of the datacube, but creates a new
datacube with the interpolated data. A feature to keep the original format for datacubes
may be implemented in the future. We will explain how to bypass this potential problem,
if necessary.
It is also possible to import/export slit-spectra, without taking into account the spatial
information, or store a cut of the current selected spectral region. These options are not
for 3D data.
Spaxels and Spectra menus
These are used to open the Spaxels inspector and the Spectra inspector. They allow
also a switch between two selection modes (option Select): one based on wavelength
ranges or another based on the distribution of the spectra in the RSS. The plot options
just send the selected regions to the respective inspector.
Exercise 10: Open both the spaxels inspector and the spectra inspector. Select a wave-
length range and right-click on the mouse. Have a look to the spaxels inspector. Now,
change the selection mode from spaxels to spectra (by clicking on the tag ‘Select’ in the
Spectra menu). Select a set of spectra (two left-clicks and then one right-click on the
mouse), and look at the spectra inspector. Go back to the default setting.
The configuration menu
! ColorMap, used to change the current colour look-up table.

This allows you to change the current settings for the visualization tool:
Exercise 11: Experiment with the different look-up tables.

! Scale, to change the colour-map scaling (also known as intensity transform table).
! Spectral Representation, used to change the way the spectra are displayed in the
spectra inspector (combine display and plot, plot only, etc.).
Exercise 12: Experiment with the different spectral representations: select a group of
spaxels and display their corresponding spectra in the spectra inspector. Try the different
modes until you get used to them.
! Interpolation. This brings up a pop-up menu where it is possible to modify the
interpolation method and its parameters. E3D has five different interpolation rou-
tines implemented, which were borrowed from PLPLOT (with permission of the
authors under the condition of not copyrighting the code). It is possible to switch
between these different interpolation methods by pulling down the FUNCTION menu in
the interpolation window. The values of the Grid Option parameter and the output
pixel size may be overwritten. Some of the routines require an additional param-
eter, which require changes from function to function (for their exact meaning, see
the E3D documentation).
Exercise 13: Import the NAME RSS.fits file in E3D, select a given wavelength range
and display the results in the spaxels inspector. Select the map option in the Shape
menu to display an interpolated map of the slice. Change the interpolation parameters
and methods to see the effect on the interpolated map. As far as we know, the optimal
interpolation is the Natural Neighbour method. However, it is a slow one. Choose a very
small value for the Grid Option, e.g. 1e-9. The smaller it is, the better it is; however,
it is also the slower.
! Spaxel Selection: This brings up a pop-up window with a menu that allows you
to change the method of selecting spaxels, switching from single to aperture. The
default method is single, and allows for the selection of spaxels one by one in
the spaxels inspector. The aperture method allows you to select all the spaxels
within a given radius (selected in the pop-up window).
There is another selection method, the pseudo-slit method. By hitting the s
key over the spaxels inspector, it is possible to switch between the single (or
aperture) selection and the pseudo-slit selection. When this mode is selected, you
just need to select two points on the spaxels inspector display (with a left-click
on the mouse) and all the spaxels over the straight line joining these two points will
be selected. As in the previous method, you can view the corresponding spectra on
the spectra inspector display by right-clicking on the mouse. The pseudo-slit
selection is useful for comparing your data with slit spectrum.
Exercise 14: Experiment with the various spaxel selection methods. Select spectra one
by one, switch to aperture selection, and to the pseudo-slit selection. See the differences
in the spectra inspector and experiment until you get used to the different methods.
The DAR tools menu

E3D includes some tools for correcting for the differential atmospheric refraction. These
! tools for determining the shift and angle of the image along the wavelength;
tools can be divided into two groups:
! tools to correct for this shift.

The first tool computes the theoretical DAR correction (Filippenko, 1982). It relies on
HEADER keywords not included in the standard Euro3D format (see the E3D user
guide), and in a proper orientation of the position table in the sky (east left, north
up). This tool is designed as the Th.Det.DAR in the DAR menu. Two other empirical
determinations of the DAR have been included. These tools rely on the presence of a
clear peak in the spatial distribution at any wavelength (e.g. a point-like source). They
are designed as PM.Det.DAR, which determines the centroid of the image based on the
principal momentum, and peak.Det.DAR, which determines the peak of the emission by
a simple maximum search.
Once the DAR correction is determined, it is possible to perform the correction, choos-
ing the Correct DAR option. It is important to understand what the program is doing
when performing a DAR correction: first of all, at each wavelength, the data are inter-
polated (using the interpolation method set in the configuration), then the intensity
corresponding to the original spaxels grid, at the new position, is determined using the
off-set of the centroid computed previously (for more detail, see for example Arribas
et al., 1999).
Any DAR correction implies an interpolation in the spatial domain. To resample the
data in the original shape is just a mathematical trick, since the recovered data come
through an interpolation. We therefore do not recommend using the DAR tools included
in E3D for science applications; however, it can be useful for a quick check of the data
and the DAR effects. We will explain below how to better correct for DAR.
Exercise 15: Experiment with DAR correction (dir: tut3). Run E3D and
import the cube star no dar.cube.fits. Open the spaxels inspector and, using the
movie effect, check the DAR effect. Click on the peak correction on the main window
menu bar (DAR tools) and perform the correction. Store the result in another file (e.g.
star dar.cube.e3d). Note that the correction is not optimal.
The command-line entry

E3D is based on a Tcl/Tk interpreter embedded with the E3D commands (see the E3D
user guide). The GUI is no more than a front end to that set of commands. It is therefore
possible to invoke any of these commands by entering them directly in the command line
entry, at the bottom of the main window.
It is also possible to run a script stored on disk by entering source scriptfile at
the command line prompt, where scriptfile is the name of the file.
Moreover, each time you execute tk e3d.tcl, the program creates a log file to record
the various actions you have performed. This log is stored in the file e3d.log. You can
therefore cut and paste from this file to run a previous set of commands.
Exercise 16: Experiment with the command line entry. Run E3D, and import
the cube VIMOS.cube.fits. Tune the maximum and minimum values to see the data.
Enter the following command lines:
recolor image heat 0
recolor image iraf 0
recolor image rainbow 0
See the effect on the display.
Exercise 17: Experiment with the scripts (dir: tut3). Run E3D and enter the
following words on the command line: source script.tcl. See what happens. Open the
file script.tcl with your favourite text editor (e.g. vi, emacs), and have a look to the
sequence of commands you have executed.
7.5.4 Menu options of the Spaxels Inspector
The top-level menu of the spaxels inspector gathers the options that affect the spaxels-
oriented representation and selection of data.
The File menu

This handles the various input and output of data (or plots) driven by the spatial distri-
bution. The data files can be slices (i.e. collections of coordinates with their corresponding
fluxes) or maps (2D images). The slices may be stored in FITS or ASCII tables (four-
column tables, with id, x, y and the flux), while the maps can be stored in FITS images.
The maps are derived from the currently displayed dataset, interpolated according to the
current configuration settings. A particular case is the save object option, that we will
explain below.
The plots are Postscripts or GIF hardcopies of the current display.
The Shape menu
This allows the user to change the way the data are represented in the spaxels inspector,
including methods combining the original spaxels shape and the image of interpolated
data (i.e. the map).
Exercise 18: Import the file INTEGRAL SB2.rss.fits, and the position table
INTEGRAL SB2 pt 219.txt (dir:tut4). Open the spaxels inspector and browse different
wavelength ranges. Switch between the different options in the Shape menu to experiment
with the various representations. Store the plots/slices obtained using the different out-
put options in the File menu. In the case of 2D FITS images, visualize them using an
external tool, e.g. DS9. View the ASCII tables using an editor, or use the Unix command
more to look at them. For the plots, display them with ghostview or xv.
The Select menu

First, this allows you to clear the current selection and start a new one using the Clear
! plot spectra: This has the same behaviour as right-clicking on the mouse.
option, but you may also:
! plot profile: This plots the intensity profile of the slice currently displayed along
the selected spaxels. It is particularly useful when combined with the pseudo-slit
selection.
Exercise 19: Hit the s key on top of the spaxels inspector display. Look at the infor-
mation labelled displayed on top of it. It should say Pseudo-Slit Selection. Select a
cut that goes through the centre of the image, and select the plot profile option. The
intensity profile along the pseudo slit will appear in the spectra inspector, fitted with a
Gaussian.
! create object: An object stands for a collection of selected spaxels. Once selected,
a new entry will appear on the Objects list box in the left-hand menu. A maximum
of nine different objects can be created; if this number is exceeded, the objects are
overwritten. Objects are the best way to store your previous selections. To get back
to any object, just click on it in the Objects list box. The spectra corresponding
to the selection of spaxels stored in this object will be displayed in the spectra
inspector.
Exercise 20: Play with the object creation (dir: tut4). On the spaxels inspec-
tor select the central spaxels and store the selection as an object. Clear the selection
and repeat the process with the outer part of the central square, and with the external
circle of spaxels. Open the spectral inspector and select alternatively each of the objects.
Look at the different selections of spectra displayed.
It is possible to store a selected group of spaxels (or object) in a separate Euro3D file,
using the save object option. Just click on one of your objects (or select a group of
spaxels) and select this menu option. If you want to see what the result looks like, you
! reverse object: this option is used to reverse the spaxels selection stored in an
can load it in the main window.
object (i.e. select the ones that were not selected, and vice versa). Click this option
and then on one of your objects.
Exercise 21: Separate sky and objects fibres (dir: tut4). Import the INTEGRAL
SB2.rss.fits file, and display a slice in a given wavelength range using the spaxels
inspector. One by one select all the spaxels in the outer ring, i.e. those corresponding to
the sky fibres. Once selected, store the selection as an object twice. Reverse the selection
of one of these objects, so that you have the central square selected. Apply the two avail-
able selections and store the results with the save object option in two different files:
obj.e3d and sky.e3d. Load these files in E3D to see the result.
Exercise 22: Mask regions creating objects (dir: tut3). Import the VIMOS.
cube.fits file in E3D. Open the spaxels inspector. Select a given wavelength range and
display the corresponding slice on the spaxels inspector. You may see four bad regions
corresponding to dead fibres. Select them and store the selection as an object. Reverse
the selection and store the object as good data.e3d. Load the newly created Euro3D file
and display the result in the spaxels inspector. You have removed the bad data.
7.5.5 The Tools menu
! Blink: Allows blinking between slices stored in the eight-length small display buffer.
This gathers the basic tools that operate over the slices:
Exercise 23: Play with the blinking facility (dir: tut4). Import the INTEGRAL
SB2.rss.fits file and display different slices at various wavelength ranges (hint: select
continuum dominated regions and emission line regions). Then, click on the blink button
and select, by left-clicking the mouse on the small displays, the slices you want to blink
in between. Click again on the blink button to start the blinking, and once again when
you want to stop it.
! Arithmetic: This allows the user to perform simple arithmetic computations
between two buffered slices. It brings up a small pop-up window whose parame-
ters are: the two slices to be processed (using their buffer number) and the number
of the buffer in which to store the result. You select the type of the computation by
pulling down the Arithm menu at the bottom of the window.
Exercise 24: Map of emission lines, continuum subtracted (dir: tut4). Load
the obj.fits (created before), and display two slices each corresponding to the wave-
length region of [OIII] and of a continuum region nearby. Subtract the continuum to the
[OIII] slice. This will be a raw estimation of the continuum-free emission-line map.
Keyboard shortcuts
! The ‘s’ key switches from the pseudo-slit selection mode to the spaxel-to-spaxel (or
There are a few keys that may be used for different operations:
! The ‘z’ and ‘u’ keys allow the user to zoom over the image. Hit the z key, with the
aperture), selection.
mouse cursor located on the upper corner of the area to zoom and once more with the
mouse cursor over the lower corner. A zoom over the selected region will be displayed
in the spaxels inspector. To unzoom, hit the u key over the spaxels inspector to go
back to the display of the whole image.
Exercise 25: Play with the zooming capability (dir: tut3). Import the VIMOS.
cube.fits file in E3D. Open the spaxels inspector. Select a given wavelength range and
display the corresponding slice on the spaxels inspector. Zoom in and out over different
regions until you get used of it.
7.5.6 Menu options of the Spectra Inspector

The top menu of the spectra inspector gathers the options related to the spectra-
oriented representation and selection of data.
The File menu

This handles the capabilities of saving spectra or hardcopy of the plots. Hardcopies may
be saved in Postscript or GIF, spectra as 2D RSS images. A single spectrum (like the
average spectrum) may be saved using FITS or ASCII tables. In E3D, a spectrum saved
as a table is formatted as follows: the file contains three columns, which are respectively
the Id (i.e. the spectral index), the wavelength and the flux. Please remember that the
stored spectrum is the mean of the currently selected (and plotted) spectra, not the sum
of them.
Exercise 26: Testing the saving capabilities of the spectra inspector (dir:
tut4). Import the INTEGRAL SB2.rss.fits file and display a slice using the spaxels
inspector. Select different groups of spaxels and display the corresponding spectra in the
spectra inspector. Save the selected spectra as a Postscript and a GIF file (hard-
copies), then as a 2D RSS image, a FITS table and an ASCII table. Visualize the outputs
using external tools, such as ghostview, xview, iraf or DS9. Repeat the operation until you
get used to it.
The Spectra menu
This menu handles the storage in memory (or buffering) of average spectra. The capa-
bility is similar to the object creation in the spaxels inspector (see above). These
spectra may be plotted, one-by-one or simultaneously, or may be used to perform basic
computations, like those exhibited in the spaxels inspector using slices.
Clicking on Save Spectrum stores a new spectrum in the memory (up to a maximum
of nine), and a new entry appears in the list box labelled Spectra. In order to display
any of the buffered spectra, just click on the corresponding entry in the list box.
It is possible to view several spectra on the same display by selecting the tag labelled
single/multiple plot on the pull-down menu, and then selecting the spectra to be
plotted using the spectra list box. Each spectrum will be plotted using a different
colour in order to distinguish them. Click on the Clear Multiple selection option to
clear the screen and create a new selection.
Exercise 27: Test the spectra buffering capability (dir: tut4). Import the
INTEGRAL SB2.rss.fits file and display a slice using the spaxels inspector. Select dif-
ferent groups of spaxels and plot the corresponding spectra in the spectra inspector.
Store each of these spectra in a separate slot. Select each of the spectra to view them in
the spectra inspector display. Change from single to multiple selection and repeat the
selection to see the difference. Store them as Postscripts and/or GIF files, and visualize
them using external tools.
! Arithmetic allows you to perform basic arithmetic computations using buffered

Other menu options are:
spectra. A window pops up (similar to the slice arithmetic one), requesting the
selection of two spectra (from the list box) and the type of arithmetic operation to
perform (pull-down menu Arithm on the bottom of the window). The result will be
stored in the first available slot.4
Exercise 28: Get used to the spectra arithmetic capability (dir: tut4).
Import the INTEGRAL SB2.rss.fits file and display a slice using the spaxels inspector.
Select various groups of spaxels and plot the corresponding spectra in the spectra
inspector. Store each of these spectra in a separate slot. Select spectra to view them
in the display of the spectra inspector. Perform different arithmetic operations with
the selected spectra and visualize them using the single or multiple display options to
compare the results.
! Sky Subtraction allows to subtract one of the buffered spectra from the dataset.
This option is mostly used to perform a sky subtraction. To use it, just select this
option in the menu, and then select the spectrum to be subtracted (i.e. the sky
spectrum).
Exercise 29: Sky subtraction (dir: tut4). Import the INTEGRAL SB2.rss.fits file
and display a slice using the spaxels inspector. Select the spaxels in the outer ring (the
sky-fibres ones), and visualize the average sky spectrum in the spectra inspector. Store
this spectrum in one of the memory slots, and perform a sky subtraction using this spec-
trum as the sky spectrum. Once the subtraction is performed it is possible to save the
whole dataset on disk. Go back to the main window and save it as an RSS or a Euro3D
dataset. It is also possible to remove the sky spaxels by storing the sky spaxels selection as
4
There is a known bug when trying to use spectra from the first position of the slot.
an object, performing a reverse object operation, and then storing the object (e.g. as
sobj.fits for RSS format or sobj.e3d for Euro3D datacube), as described in the Exer-
cise above.
The Fitting menu

E3D includes some basic fitting routines. More sophisticated routines will be discussed
later on in a dedicated section. We will describe here the fitting routines used to derive the
FWHM of a profile (cut on a slice using the spaxels inspector). Two fitting capabilities
! Fit Single Emission line: This fits the plotted spectrum with a single emission
are provided:
line (a Gaussian function), plus a continuum (as a threshold). The four free param-
eters of the fitting are the intensity of the Gaussian, the FWHM, the central wave-
length and a constant for the continuum. The result of the fit will be overplot in
red (in contrast to the data which are plotted in white). The output of the fitting
routine is printed in the xterm.
Exercise 30: Fit an emission line (dir: tut4). Import the sobj.fits or sobj.e3d
file and select the slice whose wavelength range is centred on the [OIII] emission line, with-
out any other emission line pollution, using the spaxels inspector. Plot the corresponding
spectrum using the spectral inspector. Fit this spectrum using the Fit Single Emission
line.
! Full Automatic Kinematics Analysis: This performs a kinematics analysis using
a single emission-line fitting (as described above) per spectrum. Being an automatic
procedure, the result is not fully reliable but must be considered as a quick estimation
of the kinematics on the object. It fits each spectra in the dataset with a single
Gaussian function (the emission line) plus a constant (the continuum). Once the
fitting is done, it creates four maps using the output of the fit, i.e. a map of the
line intensity, a map of the FWHM, a map of the central wavelength and a map of
the continuum emission. The fit is performed over the currently selected wavelength
range.
Exercise 31: Full kinematics estimate (dir: tut4). Import the sobj.fits or
sobj.e3d file and select a slice centred on the [OIII] emission line, without pollution
of any other emission line, using the spaxels inspector. Plot the corresponding spectrum
using the spectral inspector. Fit this spectrum using the Fit Single Emission line.
Once this first fit is performed, click on the Full Automatic Kinematics Analysis.
The routine loops over the spectra of the dataset. Once finished, four new maps will be
created in the spaxels inspector, one for each free parameter of the model (see above).
Check visually the different maps.
7.6 Sky subtraction

The sky background correction is usually applied by subtracting a sky spectrum from the
object spectra. The sky spectrum is usually obtained by computing the mean of several
spectra belonging to sky spaxels on 3D observations.
For a proper sky background subtraction, the sky spectrum must be wavelength cal-
ibrated. Figure 7.10 shows a typical sky background spectrum in optical wavelengths.
Multitude of bright emission lines can be identified in the optical sky spectrum, being
worse in the infrared spectrum. [OH] emission lines may be observed between 6500 Å and
2.62 µm, with the brightest lines above 1.5 µm. For example, in the H band, about 70
300
OH Meinel bands
200
Intensity fv (µJyarcsec–2)
V
100
OI OE
B 5577
OI
6302 8645
U
NaD
OI
5890/6
Hg 6364
Hg NI? 5460
4858 O lines 5199
s
0
4000 5000 6000 7000 8000 9000
Wavelength (Ångstroms)
Figure 7.10. Typical sky spectrum in the optical wavelength range obtained on a moonless
night at the Observatorio del Roque de los Muchachos (ORM) on La Palma island (Spain).
Most of the emission features identified in the sky spectrum are due to airglow. The Meinel
rotation-vibration bands of [OH] dominate the sky background spectrum redward of 6500 Å.
The typical broad band filters for astronomical observations are marked (figure from Benn and
Ellison, 1998).
distinct [OH] lines can be identified which are spectrally unresolved at a resolution of
∼ 17,000. The problem of subtracting such number of sky lines is a difficult one because
the absolute and relative intensities of [OH] lines are strongly temporal and spatially
dependent. For this reason, the accuracy of the sky subtraction is much better in the
continuum than in atmospheric emission lines, and the large number of [OH] lines makes
an accurate sky background subtraction in the red difficult.
But the accuracy of the sky background also depends on instrumental features. Typi-
cal instrumental problems such as flexures in the spectrograph, aberrations in the optics
or variations in the spectral efficiency of the detector also cause bad features when sub-
tracting the sky spectrum.
Therefore, atmospheric and instrumental factors complicate such an a priori easy task
of subtracting the sky spectrum from the object spectra. Figure 7.11 illustrates troubles
found when subtracting the sky background on 3D data.
Depending on the science to be done, sometimes it is very important to subtract the
sky background with a high level of accuracy. In some cases – for example in infrared
observations – the object may have a surface brightness much fainter than the sky (for
example to study faint objects). Because IFS is usually used for observing extended
objects, most of the time it is necessary to sample the sky background far enough from the
target object (sky background exposures) using a telescope off-set. The problem is worse
when IFS is used with adaptive optics, when the point-spread function (PSF) delivered
by the telescope and the adaptive optics system should be the same in both target and
background exposures for a proper and accurate sky subtraction (for a description of
Figure 7.11. The DS9 displays viewing the same RSS image before (left) and after (right) the
sky background subtraction. On the left display, several emission lines are present. After the
sky background subtraction (on the right), some residuals remain where the brightest sky lines
were located.
different ways of removing the sky background, see Allington-Smith and Content, 1998).
Therefore, the accuracy of the sky background subtraction often depends on observing
strategies.
Software to subtract the sky background

Most of the traditional astronomical software includes tasks that allow subtraction of the
sky background. However, none of these tools was specifically designed to work with 3D
data. Therefore, a few modifications of the original packages (or scripts involving several
! IRAF (http://iraf.noao.edu/) includes a few tasks to remove the sky background

packages) are needed to subtract the sky background from IFS data.
from object spectra such as SKYTWEAK or SKYSUB. SKYTWEAK is part of

the onedspec IRAF package and subtracts sky background from 1D spectra after
tweaking the sky spectra. The IRAF specred package includes the previous SKYSUB
task. Users select a subset of spaxels from a multispec format image (RSS image)
corresponding to sky spectra. They are combined (mean, median or sum) into a
master spectrum that is subtracted from another subset of spectra defined as the
object spectra.
Other methods have been implemented for estimating and removing the sky back-
ground spectrum through IRAF tasks. For example, the SVsvdfit task – an IRAF
add-on package developed at the Smithsonian Astrophysical Observatory Telescope
Data Center – is used to calculate and remove the background spectrum from spectra
without a simultaneous measurement of the sky, using eigenvector fitting techniques.
The technique is based on the iterative subtraction of continuum estimates and eigen-
vector sky models derived from singular value decompositions of sky spectra and sky
! MIDAS (http://www.eso.org/projects/esomidas/): The command SKYFIT is the

spectra residuals (Mink and Kurtz, 2001).
MIDAS task for removing the sky spectrum from long-slit data. This program makes
a polynomial fit of the sky spectrum using two different windows above and below
the object spectrum. The final sky spectrum to be subtracted is obtained from a
combination of the spectra of these two windows.
The sky background component can be also estimated using the commands
DEFINE/SKY and EXTRACT/SKY within MIDAS. The command DEFINE/SKY
allows the user to define the off-set limits of up to two sky windows, usually on
both sides of the object spectrum. The command EXTRACT/ECHELLE performs
an average extraction of the sky background using these off-sets. The two extracted
! FIGARO (http://starlink.rl.ac.uk/): The POLYSKY task is used in FIGARO to

images can be subtracted for sky background correction.
subtract the sky from long-slit data by polynomial fitting in the spatial direction
of two defined regions on either side of an object of interest. FIGARO users can
also use the command SCNSKY, which uses the minimum median algorithm to
create a sky spectrum from an image without any lines clear of stars. FIGARO also
provides the SKYLINER task removing a sky spectrum normalized by the height of
the 5577 Å [OI] emission line.
7.7 Fitting and modelling

Fitting and modelling is often useful to model the spectra and/or the 2D distribution of
the flux (or even the entire dataset) to extract more information from the data. Here we
will discuss the peculiarities of fitting and modelling 3D data.
7.7.1 Emission-line fitting

We give here a brief description on how to fit and deblend emission lines (like ionized
gas emission lines). For this purpose, we will use our own fitting routines, handling both
RSS images and datacubes (FIT3D, Sánchez et al., 2004); however, please note that these
are still under development and distributed as they are. These routines consist of a core
coded in C and a set of perl wrappers.
fit spec back , the C-coded core fitting routine
This program fits the data with the given model computing a minimization of the reduced
χ2 and using a modified Levenberg–Marquardt algorithm. The user provides a spectrum
(ASCII table in the E3D style5 ) and a configuration file where the model to fit is described
! first, a header line with four elements:

and formatted as follows:
– two integers, the first one always left to 0 (to identify the file), and a second giving
the number of individual functions included in the model;
– two floats, the first one being the goal value for the reduced χ2 and then the goal
value for the variation of this value between two consecutive iterations. When one
! for each function:

of these two limits is reached, the fitting process stops.
– a first line indicating its name;

– nine further lines (one per parameter, nine being the maximum number of param-
! each of these five elements gives:

eters allowed in FIT3D), with five numbers each.
– a first guess of the parameter value (a float value);
5
This is a three-column ASCII file containing: ID, WAVELENGTH and FLUX.
– an integer flag indicating if this parameter is to be fitted (= 1) or if it will stay
fixed (= 0);
– the lower bound for the parameter;
– the upper bound for the parameter;
– an integer flag indicating if this parameter is free (= −1) or if it is linked to the
same parameter of another function of the model (= to the number of the function
! The linking method allows the fitting of lines that share some properties (e.g. lines
in the configuration file to which it is linked, 1 being the first one).
that are kinematically coupled) or the inclusion of lines for which the line ratio is
known (e.g. the line ratio between [OIII] 5007 Å and [OIII] 4959 Å). There are two
ways to link parameters:
– additive: the parameter of this function is equal to the parameter of the function
indicated by the element number (5) plus a constant. In this case element number
(3) indicates the constant to be added and element number (4) must be set to 0.
– multiplicative: the parameter of this function is equal to the parameter of the
function indicated by the element number (5) multiplied by a constant. In this
case element number (3) indicates the multiplicative factor, and element number
(4) must be set to 1.
Although many functions are described with fewer than nine parameters, the nine lines
are mandatory (but possibly filled with zeroes). The currently available functions in
! gauss1d one-dimensional Gaussian, whose parameters are (in the following order):
FIT3D, to build a model for Observatoire de Lyon, are:
(1) the central wavelength, (2) the peak intensity and (3) the σ of the Gaussian
! eline a Gaussian emission line, whose parameters are (in the following order): (1)
function (describing the width, FWHM = 2.345 σ).
the central wavelength, (2) the integrated flux, (3) the σ of the Gaussian function
! poly1d a n-order polynomial function (n < 9), whose parameters are the coefficients
(FWHM = 2.345 σ) and (4) the velocity in km s−1 .
! back a spectrum background, included in the list of files BACKGROUND LIST (one file
of the polynomial function.
per line), whose parameters are: (1) order of the background file in the list, (2) the
intensity scaling factor, (3) the σ of a Gaussian function to convolve the background
spectra to match the resolutions and/or the velocity dispersions, and (4) the velocity
in km s−1 to shift the spectrum.
! two emission lines;

The following example illustrates a model based on four functions:
! a zero-order polynomial function;

! a background spectrum.
! The lines and the background are kinematically coupled, defining a system (see
The constraints included are as follows:
! Two lines have the flux intensity linked.

below).
As shown in this example, one of the powerful features of this tool is the possibility of
defining systems. A system is a set of emission lines which share the same velocity and
velocity dispersion. The dispersion is included as a dispersion in terms of wavelength and
not of velocity, which may be a problem for large wavelength ranges. This co-moving
set of emission lines may be fundamental for deblending several components in complex
systems. The continuum can also be coupled kinematically (or not) to a system.
For example:
0 4 0.2 0.001
...
eline
poly1d
5007 1 0 0 -1
0.01 1 -10 10 -1
1.0 1 0.2 20 -1
0 0 0 0 -1
4.6 1 0 0 -1
0 0 0 0 -1
12000 1 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
eline
back
4959 1 0 0 -1
1 0 0 0 -1
0.33 1 0.33 1 1
10.0 1 0.2 200 -1
4.6 1 0 0 1
4.6 1 0 0 1
12000 1 0 0 1
12000 1 0 0 1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
...
The result of the fitting process is displayed on the screen and stored in the following
! out.fit spectra, an ASCII file including a header line with the number of fitted
output files:
functions (i.e. the number of function included in the model), followed by a line per
function. Each line contains 19 values:
– the label of the model, as described above;
– the values of its nine parameters;
! out config.fit spectra, a configuration file where the output values of the fit are
– the nine corresponding error estimates.
used as first-guess values. This file may be useful when fitting iteratively similar
! out mod res.fit spectra, a five-column ASCII table. The columns are:
spectra with the same model.
– the wavelength;
– the original flux at this wavelength;
– the derived model;
– the residual; and
– the residual continuum.
The syntax is as follows:
fit spectra SPECTRUM.TXT CONFIG FILE [BACKGROUND LIST NBACK]
fit spec back.pl , a perl wrapper to create configuration files
This program creates the configuration file needed by fit spec back. It requires an
input spectrum in the ASCII format (see above), an ASCII file listing the emission lines
to be fitted (see tut4/emission lines.txt), the wavelength range to be fitted, a file
containing the list of backgrounds to use (if not applicable, type none), the number of
backgrounds files, and a file defining the wavelength bounds of the spectral regions to
mask.
The emission lines file is an ASCII file with two columns, one for the wavelength of the
emission line, and the other one for the name (or label) of the line; for example:
5006.84 [OIII]5007
4958.91 [OIII]4959
4861.32 Hbeta
The file defining the spectral regions to mask is a two-column ASCII file with respectively
the lower bound (minimum wavelength) and the upper bound (maximum wavelength) of
the region to mask. The syntax is:
fit spec back.pl SPECTRUM.TXT LINE FILE START WAVELENGTH
END WAVELENGTH BACK LIST NBACK MASK LIST [CONFIG]
Exercise 32: Fitting and deblending emission lines (dir: tut4). Using E3D,
import the sobj.fits or sobj.e3d file and display a slice with a wavelength range
centred in the [OIII] emission line, avoiding other emission lines. Plot the corresponding
spectrum. Save this spectrum as an ASCII table (e.g. as spec cen.txt). Exit from E3D.
Enter on the shell command-line:
fit spec back.pl spec cen.txt emission lines.txt 4800 5300 none 0 none
Then answer the following questions:
Number of systems to include:2
Redshift of the system number 0:0.03
Systemic Velocity=8993.77374 km/s
Flexibility of the velocity (km/s):500
GAUSSIAN sigma of the line (GAUSS,min,max): 3 0.1 40
Redshift of the system number 1:0.03
Systemic Velocity=8993.77374 km/s
Flexibility of the velocity (km/s):500
GAUSSIAN sigma of the line (GAUSS,min,max): 40 0.1 1000
HeII 4686.0 at 4826.58 found
Include? (y/n):n
[OIII]5007 5006.84 at 5157.0452 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 5000
Included
[OIII]4959 4958.91 at 5107.6773 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 5000
Included
Hbeta 4861.32 at 5007.1596 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 500
Included
[FeII] 4889.62 at 5036.3086 found
Include? (y/n):n
[FeII] 4905.34 at 5052.5002 found
Include? (y/n):n
[FeII] 5111.6299 at 5264.978797 found
Include? (y/n):n
System 0 contains 3 lines
HeII 4686.0 at 4826.58 found
Include? (y/n):n
[OIII]5007 5006.84 at 5157.0452 found
Include? (y/n):n
[OIII]4959 4958.91 at 5107.6773 found
Include? (y/n):n
Hbeta 4861.32 at 5007.1596 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 5000
Included
[FeII] 4889.62 at 5036.3086 found
Include? (y/n):n
[FeII] 4905.34 at 5052.5002 found
Include? (y/n):n
[FeII] 5111.6299 at 5264.978797 found
Include? (y/n):n
The data are plotted with white lines, the model in blue and the model continuum sub-
tracted in red. The green line shows the residuals. The first fit is very bad since we have
not included any background. We will now model the background as a polynomial function.
Enter the following:
cp tmp.config fit system.config
emacs fit system.config &
The file fit system.config is the configuration file used for the model. It included a
header line and the different parameters for each function included in the model (see the
file README fitting tool.txt or Section 7.1.1). We are going to add a new function to
the model, changing the header line from:
0 4 0.2 0.001
to
0 5 0.2 0.001
and adding the description of the new parameters at the end of the file:
0 4 0.2 0.001
poly1d
100 1 -1e12 1e12 -1
0.1 1 -1e12 1e12 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
Figure 7.12. Example of the use of FIT3D: snapshot of an iteration of fit spec back where a
quasi-stellar object (QSO) spectrum is fitted with three emission lines plus a continuum. The
lines show the input data, overplotting the fitted model. The dot line is the continuum. The
residuals is the line close to zero.
Again, execute the fitting procedure using this new config file as input:
fit spec back.pl spec cen.txt emission lines.txt 4800 5300 none 0 none
fit system.config
The output of the fitting process is illustrated in Figure 7.12. Play with the parameters,
or include a new system to describe the broad line until you get an acceptable fit.
7.7.2 Gas kinematics

Gas kinematics is derived from the individual fits (one or several systems) of all the
spectra of the dataset. For each spaxel (i.e. one row of the position table), the user is
provided with the velocity and velocity dispersion of the systems, the intensity of each
line of the system, the continuum level and the estimate of the error for each parameter.
In the case of grid-distributed positions (i.e. datacubes), the output of the kinematics
analysis would be a set of 2D images (or maps), one per parameter. In other cases, it
may be required to interpolate the outputs of the process to create 2D images, but it is
also possible to handle ASCII of FITS table as slices, as stated above.
Different tools are made available for gas kinematics. FIT3D includes all the subroutines
required to handle the problem in a consistent and simple way. We describe some of these
tools here.
Fitting the spectra: kin back rss.pl and kin back cube.pl
These two programs perform one fit per spectrum respectively for an RSS and a datacube,
using the model described in the configuration file. The fitting is interactive, using the
previously described fit spec back routine. The output of this fitting process is stored
in an ASCII file, which consists of the concatenation of each out.fit spectra log file
produced by the routine. In addition, a RSS (or a datacube) of the original dataset
(i.e. the trimmed section fitted), the derived model and the residual are stored in three
different FITS files, i.e.:
kin back rss org.fits,
kin back rss mod.fits and
kin back rss res.fits
or
kin back cube org.fits,
kin back cube mod.fits and
kin back cube res.fits.
The two programs require as input a 3D dataset (RSS or datacube format), a
fit spec back configuration file (as described before). Please run fit spec back.pl
! a typical spectrum from the dataset;

to create this file, using:
! an ASCII file containing the list of background spectra and the number of them (if
! the wavelength range to fit;

no background spectrum is required, just set nback to 0);
! the output filename (ASCII format); and

! a PGPLOT device for plotting the output.
The syntax for calling these tools from the command-line is as follows:
kin back rss.pl rss.fits config file back file nback
start wavelength end wavelength out file DEVICE
or
kin back cube.pl cube.fits config file back file nback
start wavelength end wavelength out file DEVICE
Creating maps: mapgrid back.pl and map interp back.pl

These programs create flux intensity, velocity and velocity dispersion maps, and their
corresponding error maps, using the outputs of the previous kinematic analysis programs.
The maps are stored in 2D FITS files, reflecting the original disposition of the spaxels
in the sky for the first one (a case of a regular-grid datacube) or interpolating the data
for the second program (for RSS or non-regular gridded IFU). They require as input the
position table of the dataset (or an ASCII slice obtained with E3D; see Section 7.5), the
files output from the previous step (kinematics analysis programs) and a label to prefix
the output FITS files. In the case of map interp back.pl, three additional parameters
! the interpolated function to be used (as a number from 1 to 5);

are required to define the kind of interpolation to perform:
! an additional parameter required by some interpolation functions;

! the final pixel size.
These three parameters are described in Section 7.5.3, and in the E3D user guide.
Based on our experience, we recommend use of the natural neighbour interpolation,
INTERPO MODE=3, with a GRID FUNC=1e-12 and a pixel size of 1/3 of the original spaxel
size.
The names of the output maps indicate their contents. Their filename starts with the
! flux, for the flux intensity;

given prefix, followed by four possible labels:
! vel, for the velocity;

! disp, for the velocity dispersion (both of them in km s−1 );
! cont for the background continuum.
The filenames end with a running number (starting from 0) indicating the emission line
that they refer to. Each of these maps has its corresponding error map, which includes
an e in the name before the label type; for example, out sobj kin flux 00.fits and
out sobj kin eflux 00.fits are the flux intensity map and its corresponding error
map for the first emission line analysed in a kinematic study, using out sobj kin as
prefix.
The syntax for calling these tools from the command-line is the following:
mapgrid back.pl slice.txt/POS TABLE out.model PREFIX OUT
or
map interp back.pl slice.txt/POS TABLE out.model PREFIX OUT
INTERP MODEGRID FUNC GRID PIX
Exercise 33: Kinematics analysis. Using E3D, import the sobj.fits or sobj.e3d
file. Select the wavelength range between 4800–5300 Å (to speed up the process) and store
it as a RSS file (sobj rss.fits). We will use the same configuration file as for the
previous exercise. It is highly recommended to fit a single spectrum before performing the
whole semi-automatic analysis. Execute the program kin back rss.pl:
kin back rss.pl sobj.rss.fits fit system tunned.config null 0
4800 5300 out sobj kin.txt /xs
This program will fit all spectra of the dataset, one by one, using fit spec back, and the
configuration file tmp.config. Each time a fit is performed it is possible to execute the
! s to store the results and jump to the next spectrum;

following options:
! a to perform a non-interactive automatic fit over the whole dataset;

! r to repeat the fit.
Since the program opens an emacs terminal with the current configuration file,6 it is
possible to modify it and repeat the fit until a suitable result is obtained. The program
stores the results from each individual fit in a log file (out sobj kin.txt in the example),
and the fitted model and residual spectra in the FITS files kin back rss mod.fits and
kin back rss res.fits.
Perform an automatic fit using a option to speed up the exercise. It is possible to check
the results with IRAF and ds9 (or a similar tool), using the commands:
disp kin back rss org.fits 1
disp kin back rss mod.fits 1
disp kin back rss res.fits 1
6
If you are not an emacs user or you do not have it installed in your system, please edit
kin back rss.pl and change the preferences to use your favourite editor.
Figure 7.13. Example of the kinematics analysis. The top-left panel shows the derived model
using the kinematics analysis in Exercise 34. The top-right panel shows the original RSS dataset,
and the bottom-left panel the residuals after subtracting the model.
as illustrated in Figure 7.13. As explained above, it is possible to map the fitted parame-
ters, creating 2D images and interpolating data if required (in case of RSS data). To do
so, run the following:
map interp back.pl sobj.pt.txt out sobj kin.txt out sobj kin 3
1e-12 1
Once the maps are created, it is possible to visualize them using IRAF:
disp out sobj kin flux 00.fits 1
disp out sobj kin vel 00.fits 1
The results are shown in Figure 7.14.
In case of velocity dispersion, it is necessary to subtract, to the observed values, the instru-
mental component. The instrumental component consists of all the dispersions in the lines
that are not physically associated with the studied source but due to the instrument,
night conditions or any other effect. The standard procedure is to perform a kinematic
analysis over a nearby emission line (over the non-sky-subtracted data), and to subtract
the obtained velocity dispersion to the observed object (with a quadratic subtraction).
It is important to note here that the velocity dispersion is fitted assuming a dispersion
in Å, that it is not the proper approach for wide wavelength ranges. Kinematics analyses
are normally performed over a narrow wavelength range, centred on the emission lines of
interest; this is why the current approach is used. It is also easier to decontaminate this
method from instrumental effects in the velocity dispersion.
Figure 7.14. Example of the use of FIT3D. It shows the flux intensity map (left panel) together
with a velocity map (right panel) derived with the automatic line fitting performed in Exercise 34.
7.8 Decoupling spectra of different objects: crowded

field spectroscopy
3D data can be seen as a collection of narrow-band or monochromatic images of the same
object at different wavelengths. Assuming a smooth variation in the morphology of the
objects with wavelength, it is possible, using 3D data, to apply techniques of decoupling
or deblending, normally used in the analysis of 2D imaging, and to extend them to
3D (Garcı́a-Lorenzo et al., 2005; Sánchez et al., 2006a). This method is particularly
! when we have low signal-to-noise images of objects, all with the same morphology
interesting:
! when we have a crowded field where we want to deblend the spectra (e.g. the core
in the image (e.g. QSO lenses; Wisotzki et al., 2003);
! when we want to deblend two well-defined components on the same object (e.g. QSO-
of globular clusters or galaxy clusters; Sánchez et al., 2006b); or
hosts or bulge-to-disk decomposition, or point-like sources over bright backgrounds,

like in XPNs, Sánchez et al., 2004).
There are different implementations of this technique in the literature (Roth et al.,
2004; Sánchez et al., 2004; Wisotski et al., 2004; Fabrika et al., 2005; Garcı́a-Lorenzo
et al., 2005; Sánchez et al., 2006a, 2006b). We will not describe all of them, but take
a particular case, namely the extension of galfit to 3D, or galfit3d (Sánchez et al.,
2006a, 2006b). The galfit (Peng et al., 2002) program is a program for modelling 2D
images of one or several objects using analytical models for galaxies and/or point-like
sources, convolving them with a PSF.
In the case of galfit3d, we have extended the technique to model regular-gridded
3D data (i.e. datacubes), which may be interpolated versions of non-regular ones (e.g.
Garcı́a-Lorenzo et al., 2005), and decouple the spectra of different objects in the field
and/or deblend the spectra from different components of the same object (Sánchez et al.,
2006a, 2006b). A major advantage of IFS data is that it is possible to impose links over the
parameters used for each n-fits (one per wavelength), on the basis of a smooth variation
of the morphological parameters, or we can force them to a particular value if we have
a source of external information (e.g. high-resolution Hubble Space Telescope images;
Garcı́a-Lorenzo et al., 2005; Sánchez et al., 2006b). In both cases the fitting process is
iterative, with a first loop of non-acceptable results, and a set of iterations where the
morphological parameters are constrained to different values, using as input parameters
the output of the previous fits (Sánchez et al., 2006a).
7.8.1 Components of galfit3d

galfit3d is a collection of perl scripts, including:
galfit3d.pl
This program performs a modelling of an input datacube creating, for each wavelength,
a narrow-band image of the width of a spectral pixel, and fitting it using galfit (Peng
et al., 2002). The output of the fitting is stored in a log file, which consists of the
concatenation of the different outputs (e.g. galfit.01) produced by galfit for each
wavelength. A quality log file is also stored, with an integer flag that indicates if the
fitting process has finished the minimization (1) or ended in abnormal break (0). The
syntax is:
galfit3d.pl INPUT CUBE.fits MOD CUBE.fits RES CUBE.fits
GALFIT CONF LOGFILE QUALITY.txt
GALFIT CONF: A) object.fits, B) output.fits
galfit3d force.pl
This program works like galfit3d.pl, but using as guessed values for the fit the output
values of a previous iteration (INPUT LOGFILE). Each parameter can be smoothed using a
polynomial function, or forced to be a given value. This is indicated with the INPUT NPOLY
! the number of the model in the galfit configuration file whose parameter needs to
configuration file, which is a three-column ASCII file whose elements per line are:
! the number of the parameter of this model to be smoothed;

be smoothed;
! the order of the polynomial function to fit (if value is an integer) or fixed value (if
value is a float).
The syntax is:
galfit3d force.pl INPUT CUBE.fits MOD CUBE.fits RES CUBE.fits
GALFIT CONF INPUT LOGFILE INPUT NPOLY OUTPUT LOGFILE QUALITY.txt [PLOT]
extract par g3d.pl

This program allows the extraction of the fitted values for a parameter of a given model
along the spectral pixels and the storage of them in an ASCII output file, plotting this
parameter along the spectral pixels using PGPLOT with the defined device. It requires as
input the output log files obtained with galfit3d.pl or galfit3d force.pl, the name
of the output file where the results are to be stored, the number of the model and the
number of the parameter to extract. The syntax is:
extract par g3d.pl INPUT LOGFILE QUALITY.txt OUTPUT FILE
model number par number DEVICE [min max]
extract spec g3d.pl

This program allows the extraction of the integrated spectrum corresponding to an
object (or component of an object) modelled with a 2D model, and the storage of it
as ASCII spectrum (see Section 7.5). It requires as input the output log files obtained
with galfit3d.pl or galfit3d force.pl, the name of the output file where the results
are to be stored, the number of the model in the galfit configuration file, the zero point
of the magnitude indicated in that configuration file, the starting wavelength and the
step between wavelengths for the fitted datacube. The syntax is:
extract spec g3d.pl INPUT LOGFILE QUALITY.txt OUTPUT FILE
model number mag zeropoint CRVAL CDELT
7.8.2 Installation of galfit3d

galfit3d is still under development. We distribute in these proceedings a beta version
as it is (no warranty). A large number of simulations is still needed before producing
reasonable outputs for the most usual cases. It requires having galfit and IRAF installed
on your computer. We recommend that the user modifies the source code of galfit in
order to increase the accuracy of the produced output, not to be truncated to the second
decimal of magnitude (see galfit documentation). As a set of perl scripts, it requires
the modules described in Section 7.4 to be installed too. Once these modules are installed,
please update the following lines in the perl scripts with your own paths:
require /data/sanchez/code/R3D/my.pl;
$galfit="/data/sanchez/galfit/galfit";
$cl="/usr/local/bin/cl";
Exercise 34: 3D modelling. It is beyond the scope of this tutorial to make a complete
description of the deblending and modelling options for 3D spectroscopy. We will therefore
perform a very simple modelling on GMOS data. The data presented here are those of
a nearby QSO, where we would like to deblend the bright QSO from the extended Host
galaxy. To do so, the model mimics the point-like source using a Gaussian convolved with
a PSF, and the Host galaxy using a Sersic profile, with the Sersic index equal to 1, i.e. a
Disk galaxy. We will guide the fit by forcing a scale-length, ellipticity and position angle
for the galaxy, and we will not perform a second iteration. Run the program:
galfit3d.pl input.cube.fits input cube.mod.fits
input cube.res.fits galfit.qso fix input cube.fit.log
input cube.fit.quality
The program will loop over all the slices in the dataset and fit them with the given model.
The output of the fit is stored in two files: input cube.fit.log (the log file), with the
direct outputs from galfit, and input cube.fit.quality (a file containing a flag indi-
cating if the fit ended normally or crashed). It also produces two datacubes: one with the
fitted model and another with the residuals from the fitting technique. It is possible to
inspect the progress of the fit, and visualize any of the fitted parameters, by executing:
extract par g3d.pl input cube.fit.log input cube.fit.quality
junk.junk 1 3 /xs 22 18
It is also possible to visualize each of the individual fits using IRAF, by typing:
cl < display.cl
Once the fit is finished, it is possible to extract the deblended spectra of each components
by executing:
extract spec g3d.pl input cube.fit.log input cube.fit.quality
spec nuc.txt 1 25 6600 2.3
extract spec g3d.pl input cube.fit.log input cube.fit.quality
spec host.txt 2 25 6600 2.3
7.9 DAR correction using external tools
As mentioned before, E3D contains its own tools to perform the DAR correction. However,
in most cases, they cannot produce results accurate enough, and are only valid for a quick
correction. Let us remember that any DAR correction requires the spatial resampling of
the data, which produces a loss of the initial spatial configuration of the spaxels. (This
is not critical, but is important to know.)
We discuss here other tools that perform empirical DAR corrections for datacubes.
‘Empirical’ means that the DAR correction is determined directly from the data, which
is a unique feature of 3D spectroscopy. For doing this you need a bright object with a
well-defined intensity peak at any wavelength. The empirical correction is determined
by looking to the spatial coordinates of the intensity peak for each wavelength and esti-
mating the shift from one to the other. In many cases the empirical correction works
better than the theoretical one, based on the Filippenko (1982) formulae. The reason is
that the theory behind these formulae assumes that the layer of the atmosphere with
equal refraction index is flat and perpendicular to the zenith of the telescope. How-
ever, it is well known that this in only a first-order approximation since these layers
depend strongly on the orography of the observatory. The next sections describe these
tools.
7.9.1 imcntr cube.pl

This program finds the centroid of an object (using a first-guess position) at each wave-
length, and smoothing the centroid position variation from wavelength to wavelength
using polynomial fittings. It stores the results in a two-column ASCII file. It requires
as input the datacube used for the analysis, the first-guess centroid of the object, the
order of the polynomial functions to fit among the x and y coordinates, and the name
of the output file. In addition, it is possible for the analysis to exclude a given number
of spectral pixels both at the beginning and at the end of the wavelength range. The
syntax is:
imcntr cube.pl input cube.fits X Y NP X NP Y centers list.txt
[N START] [N END]
7.9.2 cube shift.pl

This program applies the DAR correction previously computed, recentring an object
along each wavelength. It requires as input the names of the input and output datacubes,
the list of centroid positions, the final coordinates of the object and the name of an output
IRAF script. This script can be used further to apply the same correction on another
datacube, by modifying the relevant values. This option is interesting for datacubes where
there is no bright object to trace the centroid positions, but for which it is possible to use
other observations of a bright nearby star (a type of DAR calibration). The syntax is:
cube shift.pl input cube.fits output cube.fits centers list.txt
X0 Y0 shift.cl
7.9.3 cube shift delta.pl

This program performs a global and spatial shift of a datacube. It requires as input the
names of the input and output datacubes and the relative shifts in the x and y axis. The
syntax is:
cube shift delta.pl input cube.fits output cube.fits delta x delta y
Figure 7.15. Example of the output of the script imcntr cube.pl: the x (top) and y (bottom)
coordinates of the centroid of a star as a function of the spectral pixel numbers, for a datacube.
To use these scripts, it is first necessary to edit the files and change the line
require /data/sanchez/perl/MY/my.pl;
to fit your own path. It is also necessary to run a mkiraf before running the programs
in the directory where you have the dataset, since these programs are simply IRAF
wrappers.
Exercise 35: Empirical DAR correction (dir: tut3). Using IRAF with DS9 (or
any similar tool), display the channel 300 of a file, e.g. disp star no dar.cube.fits
[*,*,300] and determine the centroid of the star in the field, using imexam or any similar
tool. The centroid is near the coordinates (10.5,11.5). To determine the centroids along
the wavelength, run:
imcntr cube.pl star dar.cube.fits 10.5 11.5 2 2 centers list.txt
10.5 11.5 7.5 14.5
The output of this program is illustrated in Figure 7.15, with the distribution of the x
and y coordinates of the centroid of the star as a function of the spectral pixel numbers
(representing the wavelength). The list of centroids is stored in the centers list.txt
file, used as input in the correction process:
cube shift.pl star dar.cube.fits star no dar.cube.fits
centers.lis 10.5 11.5 shift.cl
The file star no dar.cube.fits contains the DAR corrected datacube.
Figure 7.16. View of a section of IFS raw data obtained with PMAS in the PPAK mode.
Each dark line corresponds to the projection of a spectrum along the dispersion axis, which, in
this case, corresponds to the X-axis. Spectra are separated by 7 pixels across the Y-axis, and
projected in a few pixels, contaminating the adjacent spectra.
7.10 Reduction of fibre-feed IFS data

7.10.1 The raw data
The raw data from fibre-feed spectrographs consist of a collection of spectra, stored as
a 2D frame and aligned along the so-called dispersion axis. This axis can be the x- or
y-axis, or neither, although it tends to be almost aligned with one of them. Figure 7.16
shows an example of IFS raw data, corresponding to PMAS (Potsdam Multiaperture
Spectrophotometer; Roth et al., 2005) in the PPAK mode (Kelz et al., 2006), illustrating
this distribution. Each spectrum is spread over a certain number of pixels along the
perpendicular direction to the dispersion (cross-dispersion axis). The shape of the profile
along the cross-dispersion direction is similar to an asymmetrical Gaussian. There is
therefore a contamination from each spectrum to the adjacent ones. This is what is well
known as the cross-talk (see Figure 7.16). The profile changes for each spectrum along the
cross-dispersion axis and, in general, is more asymmetrical on the edges of the pseudo-slit
than in the centre. It also changes along the dispersion axis. Well-designed instruments
have well-defined and narrow profiles, and a limited cross-talk. However, there is always
a compromise between the number of fibres ‘packed’ in the spectrograph, the size of
the charge coupled device (CCD) and the width of the profiles and cross-talk. As we
quoted above, the spectra are not perfectly aligned along the dispersion axis, due to the
configuration of the instrument, its setup, the optical distortions, the instrument focus
and the mechanical flexures (if any). In addition, the transmission changes from fibre-
to-fibre due to physical differences among them, differences in the tension suffered by
the fibres at different positions, and slight mis-alignments of the fibres in the pseudo-slit.
These effects also affect the width of the projected profiles across the cross-dispersion
and the dispersion axis, and introduce small differences in the spectral dispersion from
fibre-to-fibre.
7.10.2 IFU data reduction

The data reduction of all fibre-feed IFUs shares similar steps, due to the similarities of
their raw data. We assume, as a starting point for the reduction, that the data have been
bias subtracted and flat-field corrected (when possible) with an external tool. Once this
! identification of the position of the spectra on the detector for all pixels along the
pre-processing is complete, the data reduction consists of:
! extraction of each individual spectrum;

dispersion axis;
! distortion correction of the extracted spectra and determination of the wavelength
! re-arrangement of the spectra in the sky position.

solution (i.e. dispersion correction); and
Each of these reduction steps has been treated with a set of command-line routines in
R3D (Sánchez, in preparation), a package specially written for reducing fibre-feed IFS
data. We briefly describe below the different steps required to perform a reduction.
Identification of the position of the spectra on the detector

To identify the position of the spectra on the detector you need a continuum exposure (e.g.
a dome-flat), which, in the case of instruments that suffer from flexures (i.e. displacement
of the spectra in the CCD at different positions of the telescope), should be taken with
the telescope pointing to the same location as your science exposure. The location of the
spectra is found to be at a given column in the CCD by comparing the intensity at each
row along the column (or a co-added set of rows around it) with those of n adjacent
pixels, checking for those pixels that they verify the maximum criteria, using peak find.
This program requires as input the continuum raw data frame, a flag indicating the
principal direction of the dispersion axis (0 for x, 1 for y), the width around the central
column where the intensities (for low-level signal-to-noise frames) are to be co-added, a
flag indicating if the results are to be plotted (1) or not (0), the number of subframes
to plot in the screen, the number of adjacent pixels to check for the maximum criterion,
the minimum distance between adjacent spectra (projected in the cross-dispersion axis),
the fraction of the intensity peak to use as a threshold for detecting intensity peaks
(i.e. spectra), and the name of an ASCII output file containing the location along the
cross-dispersion axis of the detected peaks. The syntax is:
peak find RAW.fits Spectral axis[0/1] Coadd width plot nplot
nsearch DMIN IMIN(% of the MAX) OUTFILE [colum search]
Then, the tracing of the peak intensities along the dispersion axis is done by
looking for maxima around each original location within a given window, using
trace peaks recursive. This program requires as input the continuum raw data frame,
a flag indicating the dispersion direction (0 means along x, 1 along y), the file containing
the location of the intensity peaks (i.e. the output of the previous program), the width
of the window where the intensities along the dispersion axis are to be co-added around
each column, the width (in pixels) indicating where to look for a new centroid of the
peak along the cross-dispersion axis, a flag indicating if the results are to be plotted (1)
or not (0), the number of subframes to plot in the screen, the number of adjacent pixels
to check for the maximum criterion, and the output file, a 2D FITS file where the x-axis
corresponds to the original dispersion axis and the y-axis indicates the number of traced
spectra. The value stored in each pixel is the location of the peak of the centroid in the
original frame of the corresponding spectrum at the corresponding spectral pixel. This
file is called ‘trace’. The syntax is:
trace peaks recursive RAW.fits Spectral axis[0/1] PEAKS FILE
coadd width search width plot nplot nsearch TRACE.fits [y shift]
2nd search width]
Extraction of individual spectra
After tracing the location of the spectra on the CCD, the next reduction step is to extract,
for each spectrum, the flux corresponding to each spectral pixel along the dispersion axis.
The simplest method to perform this extraction is to co-add the flux within a certain
aperture around the ‘trace’ of the spectra in the raw data, and store it in a 2D image.
This procedure is named aperture extraction, and it has been implemented in R3D. The
x-axis of the resulting image corresponds to the original dispersion axis, while the y-axis
corresponds to the ordering of the spectra along the pseudo-slit. This is the so-called
row-stacked spectra representation or RSS (Sánchez, 2004). The extraction is performed
using extract aper.
This program requires as input a raw data frame, a flag indicating the direction of the
dispersion axis (0 for x, 1 for y), the trace frame, the aperture width for the extraction
and the name of the output file. The syntax is:
extract aper RAW.fits Spectral axis[0/1] TRACE.fits Aperture
OUTPUT.fits [SHIFT]
Distortion correction
Most of the spectrographs do not disperse the light homogeneously along the cross-
dispersion axis. The dispersion is distorted, being larger on the edges of the slit than in
the centre. This distortion is sometimes called the ‘C’ distortion, due to its shape on the
CCD. In the case of fibre-feed spectrographs, additional distortions are introduced due
to the way the fibres are placed in the pseudo-slit. Once the spectra are extracted, these
distortions are self-evident. The left panel of Figure 7.17 shows an example of extracted
spectra, corresponding to an arc exposure obtained with PMAS in the PPAK mode. The
‘C’ distortion and the discontinuities in the dispersion solution along the cross-dispersion
may be clearly seen. These discontinuities and the slight differences in the dispersion
from fibre to fibre do not allow a 2D modelling of the distortion map with an analytical
function to be created, neither at the level of the extracted spectra, nor at the level of
the raw data. The distortions have to be corrected fibre to fibre before finding a common
wavelength solution.
R3D performs this correction by a two-step procedure, using arc calibration lamp expo-
sures (like the one in Figure 7.17). First, the peak intensity of a single emission line
is traced along the cross-dispersion axis and shifted to a common reference by a linear
shift. This is done by the program dist cor. This program requires as input the aper-
ture extracted file, the output corrected file, the output first-order distortion file, a flag
indicating whether to smooth (1) or not (0) the solution along the cross-dispersion axis,
the reference pixel in the dispersion axis to look for the emission line peak, the width of
the window in pixels to look for this peak, the number of adjacent pixels to check for the
maximum criterion, and a flag indicating whether to plot the results (1) or not (0). The
syntax is:
dist cor EXTRACTED.fits CORRECTED.fits DISTORTION CORRECTION.txt
SMOOTH[0/1] start index delta index nsearch plot [n fib] [nsigma]
[center index]
After this the intensity peak of a set of selected emission lines is traced, and a poly-
nomial distortion correction is determined in order to recentre all the lines to a common
reference. This is done using the program mdist cor sp. This program requires as input
the output of the first-order correction, an aperture to look for emission lines, a nσ
threshold over the average intensity to look for emission line peaks, the order of the
Figure 7.17. Example of extracted spectra from an arc exposure obtained with PMAS in the
PPAK mode, before correcting for the distortion in the dispersion along the cross-dispersion
axis (left panel) and after (right panel). The x-axis corresponds to the dispersion axis.
polynomial function to fit the distortion correction (spectrum by spectrum), the output
file with the distortion-corrected frame, the output file containing the 2D map of the
corrections, the range of pixels in the dispersion direction to consider for the analysis,
and a flag indicating if results will be plotted (1) or not (0). The syntax is:
mdist cor sp EXTRACTED.fits APERTURE NSIGMA NPOLY OUTPUT.fits
DISTORTION.fits NX min NX max plot [NSIGMA2] [BOX] [NY CEN]
Wavelength solution
To derive the wavelength coordinate system, the emission lines of an arc exposure are
first identified using an interactive routine. The distortion-corrected spectra of the arc are
then transformed to a linear wavelength coordinate system by a 1D spline interpolation,
assuming a polynomial transformation between both coordinate systems. The required
transformation is stored in an ASCII file to be applied on the science data at a later
stage. This procedure is done using the disp cor.pl script.
This program requires as input the output of the distortion correction analysis, the
starting wavelength and the wavelength step needed for the final spectral pixels, the
aperture to look for emission lines, the row number of the central spectrum over which
to perform the analysis, width of pixels in the cross-dispersion axis for co-adding the
spectra (to increase the signal-to-noise ratio), the order of the polynomial function to fit
the wavelength solution, the output file containing the spectra transformed to a linear
wavelength coordinate system, the output file containing the dispersion or wavelength
solution applied, and a flag indicating if results will be plotted (1) or not (0). The syntax
is:
disp cor.pl EXTRACTED.fits CRVAL CDELT APERTURE NY SPECTRA
WITDH NPOLY OUTPUT.fits DISPERSION.txt plot
Additional steps
Once wavelength is calibrated, the data need to be corrected from fibre-to-fibre differen-
tial transmission, combined with other exposures (using with off-sets), corrected for the
differential atmospheric refraction, sky-subtracted and flux-calibrated. R3D includes tools
to perform some of these operations, but it is beyond the scope of this quick reference
guide to explain all of them in detail here. Moreover, they may strongly depend on the
instrument used for the acquisition.
Exercise 36: Reducing data with R3D7 (dir: tut6). We are going to perform a
simple data reduction on PMAS Roth et al. (2005) data in the PPAK mode (Kelz et al.,
2006) using R3D. First, we determine the number of spectra and their location along the
cross-dispersion axis for the central column of the CCD:
peak find.pl DomeFF Cont Galaxy.fits 0 2 1 4 2 3.5 0.10
DomeFF Cont Galaxy.peaks
Then, we trace the location of the peak intensity of the spectra along the cross-dispersion
axis of each pixel in the dispersion axis:
trace peaks recursive DomeFF Cont Galaxy.fits 0
DomeFF Cont Galaxy.peaks 2 2 1 4 1 DomeFF Cont Galaxy.trc.fits
Once the location of the spectra in the CCD (trace) has been determined it is possible to
extract them:
extract aper.pl IRAS18131 A.fits 0 DomeFF Cont Galaxy.trc.fits 5
IRAS18131 A.ms.fits
The C distortion correction is then determined and corrected:
dist cor DomeHe ThAr.ms.fits DomeHe ThAr.dc0.fits
DomeHe ThAr.dist.txt 0 662 30 2 1 2 2
mdist cor sp DomeHe ThAr.dc0.fits 3 300 3 DomeHe ThAr.dc1.fits
DomeHe ThAr.dist.fits 17 1011 1 2 5 3
The distortion correction determined using the He lamp exposure can be applied to the
science data:
mdist cor external IRAS18131 A.ms.fits DomeHe ThAr.dist.txt
DomeHe ThAr.dist.fits IRAS18131 A.dc1.fits 0 1.353
Once transformed to a common reference, it is possible to find the wavelength coordinate
solution:
disp cor.pl DomeHe ThAr.dc1.fits 3700 1.6 3 180 3 4
DomeHe ThAr.disp cor.fits DomeHe ThAr.disp.txt 1
7
Data provided by K. Kehring.
and apply it to the data:
disp cor external.pl IRAS18131 A.dc1.fits 3700 1.6
IRAS18131 A.disp cor.fits DomeHe ThAr.disp.txt 0
The differences in the fibre-to-fibre transmission can be determined using dome-flat expo-
sures, and correct the science frames from them:
fiber flat.pl mDomeFF Cont IRAS18131 A.disp cor.fits
fiber flat IRAS18131 A.fitsimarith.pl IRAS18131 A.disp cor.fits /
fiber flat IRAS18131 A.fits
IRAS18131 A.fc.fits
REFERENCES
Allington-Smith, J., Content, R. (1998), PASP, 110, 1216
Arribas, S., Mediavilla, E., Watson, F. ET AL. (1998), Fiber Optics in Astronomy III, 152,
149
(1999), A&AS, 136, 189
Becker, T. (2001), 3D-Spektroskopie hintergrundkontaminierter Einzelobjekte in Galaxien der
lokalen Gruppe, PhD Thesis, Potsdam University
Benn, C., Ellison, S. (1998), La Palma night-sky brightness, ING technical note 115
Fabrika, S., Sholukhova, O., Becker, T., Afanasiev, V., Roth, M., Sánchez, S.F.
(2005), A&A, 437, 217
Filippenko, A.V. (1982), PASP, 94, 715
Garcı́a-Lorenzo, B., Acosta-Pulido, J.A., Megias-Fernández, E. (2002), Galaxies: the
Third Dimension, 282, 501
Garcı́a-Lorenzo, B., Sánchez, S.F., Mediavilla, E. González-Serrano, J.I., Chris-
tensen, L. (2005), ApJ, 621, 146
Glazebrook ET AL. (2001), http://pdl.perl.org/
Kelz, A., Verheijen, M.A.W., Roth, M.M., Bauer, S.M., Becker, T., Paschke, J.,
Popow, E., Sánchez, S.F., Laux, U. (2006), PASP, 118, 129
AN, 325, 159
Mink, D.J., Kurtz, M.J. (2001), Astronomical Data Analysis Software and Systems X. ASPCS,
238, 491
Pearson, T. (1995), http://www.astro.caltech.edu/∼tjp/pgplot/
Pécontal-Rousset, A., Copin, Y., Ferruit, P. (2004), AN, 325, 163
Peng, C.Y., Ho, L.C., Impey, C.D., Rix, H. (2002), AJ, 124, 266
Roth, M.M., Becker, T., Kelz, A., Schmoll, J. (2004), ApJ, 603, 531
Roth, M.M., Kelz, A., Fechner, T. ET AL. (2005), PASP, 117, 620
Sánchez, S.F. (2004), AN, 325, 167
Sánchez, S.F. (2006), AN, 327, 850
Sánchez, S.F. (2008), Euro3D Visualization Tool: User Guide for E3D, v1.3.7. Research Train-
ing Network. Available online at http://www.aip.de/Euro3D/E3D/#Docu (July 2009)
Sánchez, S.F., Cardiel, N., in preparation
Sánchez S.F., Garcı́a-Lorenzo, B., Mediavilla, E. ET AL. (2004), ApJ, 615, 156
Sánchez S.F., Garcı́a-Lorenzo, B., Jahnke, K. ET AL. (2006a), NewAR, 49, 501
Sánchez, S.F., Cardiel, N., Verheijen, M., Benitez, N. (2006b), Proceedings of the ESO
and Euro3D workshop ‘Science Perspectives for 3D Spectroscopy’, Garching (Germany),
October 10–14 (2005) (astro-ph/0512550)
Walsh, J.R., Roth, M.M. (2002), Msngr, 109, 54
Wisotzki, L., Becker, T., Christensen, L., Helms, A., Jahnke, K., Kelz, A., Roth,
M.M., Sánchez, S.F. (2003), A&A, 408, 455

3D Spectroscopy in Astronomy PDF

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

3D Spectroscopy in Astronomy PDF

Caricato da

Copyright:

Formati disponibili

This page intentionally left blank

Simultaneously storing both spectral and spatial information, 3D spectroscopy oﬀers a

Volumes in this series

Cambridge University Press

© Cambridge University Press 2009

This publication is in copyright. Subject to statutory exception and to the

ISBN-13 978-0-511-76986-3 eBook (NetLibrary)

Cambridge University Press has no responsibility for the persistence or accuracy

List of contributors page viii

List of abbreviations xiii

1 Introductory review and technical approaches 1

2 Observational procedures and data reduction 40

4 Analysis of 3D data 126

6 Extragalactic studies and future integral field spectroscopy science 200

7 Tutorials: How to handle 3D spectroscopy data 225

Matthew A. Bershady, University of Wisconsin, Madison, USA

Ahumada, Andrea Verónica Observatorio Astronómico de Córdoba

2dF The Two Degree Field system

! Astrophysikalisches Institut Potsdam (Germany, coordinator)

! IoA University of Cambridge (UK)

Each data cube slice represents

... final data cube

Single or co-added spectra

Figure 1.1. Schematic representation of a datacube, produced as the result of 3D observations.

! to promote 3D technique with selected science projects

! to develop new observing and data reduction techniques

Diameter Access 3D ∆λ Maximum FoV Maximum resolution

NGST 8.0 • IFMOS proposed

Figure 1.3. The conceptually problematic picture of a ‘datacube’

! to provide feedback from observers to technology

1.2 Introductory review

Methods of image dissection: spatial sampling

Detectors and spectrographs

Figure 1.8. Rockwell 2K × 2K Infrared Array (left), 4K × 4K mosaic (right).

Figure 1.10. Diﬀerent ways of allocating spectra on a detector: a) maximizing number of

Spectral resolution and wavelength coverage

! SAURON (WHT) [Spectrographic Areal Unit for Research on Optical Nebulae

(William Herschel Telescope)], which has a wavelength coverage of 54 nm in the

! PMAS (CAHA 3.5 m) [Potsdam Multiaperture Spectrophotometer], with a wave-

length coverage of up to 680 nm in the interval of 350−900 nm. Instrumental eﬃciency

Coupling the IFU to the spectrograph

Table 1.2. Examples for 3 types of 3D spectrographs

Nspectra FoV [′′ ] scale [′′ ] R λ range/∆λ [nm]

SAURON LR 1577 41 × 33 0.94 1200 450–700/54

Examples of real IFUs

Extracting the spectra

Data representation, data formats

1.2.2 Comparison with classical techniques

! a posteriori advantage, pointing;

! absence of slit eﬀects;

A posteriori advantage, pointing

a shift in x,y as a function of wavelength in the familiar picture of a datacube; however,

Crowded field 3D spectroscopy

Ultra-deep faint object 3D spectroscopy

1.3 A brief history of 3D spectroscopy

Figure 1.23. DensePak-2 at the KPNO 4 m telescope.

Figure 1.24. INTEGRAL at the Nasmyth focus of the WHT.

SB1 SB2 SB3 Sk fibers

f20 mm f20 mm f20 mm

bserving fibers bserving fibers bserving fibers

Figure 1.25. INTEGRAL fiber bundles.

Figure 1.26. SAURON at the WHT, La Palma.

1.4 Technical approaches